<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Alex</title>
    <description>The latest articles on DEV Community by Alex (@alexar76).</description>
    <link>https://dev.to/alexar76</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3927872%2Fdf501cac-6862-47af-8b22-4e3067696337.jpeg</url>
      <title>DEV Community: Alex</title>
      <link>https://dev.to/alexar76</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/alexar76"/>
    <language>en</language>
    <item>
      <title>My AI Agent Bought a $2.75 Bus Ticket. Then It Had to Prove It Paid.</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Sun, 15 Nov 2026 06:33:17 +0000</pubDate>
      <link>https://dev.to/alexar76/a-telegram-bot-bought-my-275-bus-ticket-what-happens-if-it-retries-4c8h</link>
      <guid>https://dev.to/alexar76/a-telegram-bot-bought-my-275-bus-ticket-what-happens-if-it-retries-4c8h</guid>
      <description>&lt;p&gt;At 13:15, I approved a $2.75 bus ticket in Telegram. A minute later, my AI agent, ARGUS, sent me a PDF for Hudson Link’s H07 route from West Nyack to Tarrytown. I opened it and saw the boarding barcode. Then ARGUS reported that it still hadn’t matched the charge to the carrier’s receipt.&lt;/p&gt;

&lt;p&gt;That left an awkward question. If I ask ARGUS to check again, will it look up my ticket or buy me another one?&lt;/p&gt;

&lt;h2&gt;
  
  
  One tap to buy the ticket
&lt;/h2&gt;

&lt;p&gt;I used &lt;code&gt;/buy_hudson&lt;/code&gt; in a private chat. ARGUS asked for the passenger’s first name and surname, showed the route and total, and waited. I tapped &lt;strong&gt;Confirm purchase&lt;/strong&gt;. An isolated Chromium session filled the carrier’s checkout using Cybersource’s hosted card fields. The browser got a short-lived payment token and submitted one order. ARGUS sent the resulting ticket PDF back to the same chat. &lt;a href="https://developer.cybersource.com/demo/doc/microform_doc.html" rel="noopener noreferrer"&gt;Cybersource Microform&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;There were two payments in the experiment. The day before the ticket purchase, an operator used ARGUS’s x402 client to issue a &lt;strong&gt;$5 prepaid card&lt;/strong&gt; from Laso. That payment was &lt;strong&gt;5 USDC on Base&lt;/strong&gt;. Buying the &lt;strong&gt;$2.75 ticket&lt;/strong&gt; used that card at Hudson Link’s ordinary USD checkout. The ticket charge appears in the card activity. &lt;a href="https://docs.laso.finance/guides/usa-prepaid-card" rel="noopener noreferrer"&gt;Laso card guide&lt;/a&gt; · &lt;a href="https://base.blockscout.com/tx/0xc8bd81d9d1d39327736bc7b7a8f490658b50362aca8e295ee8cd655f20099c68" rel="noopener noreferrer"&gt;Base transaction for card issuance&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffrb9jj9uyom9u2v1pmu6.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffrb9jj9uyom9u2v1pmu6.png" alt="Mermaid flowchart showing the $2.75 Telegram approval, one card payment, and the later check of the saved order against the carrier receipt and Laso card activity." width="800" height="1323"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The checkout now saves the order number as it arrives. On the original run, the PDF reached Telegram before the receipt check was complete.&lt;/em&gt; &lt;a href="https://modelmarket.dev/blog/media/argus-hudson-link-2026-11-15/checkout-flow.mmd" rel="noopener noreferrer"&gt;Mermaid source&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The receipt was right. The amount was a string.
&lt;/h2&gt;

&lt;p&gt;Laso sent the amount back as &lt;code&gt;"2.75"&lt;/code&gt;. The first matcher skipped it because it expected a JSON number. The purchase had gone through. Our code had lost track of one piece of evidence.&lt;/p&gt;

&lt;p&gt;The fix parses either representation as integer cents before comparing it with the approved amount. This reduced example handles unsigned decimal amounts and runs on Node.js 22 with no dependencies. The production check also matches the merchant and card reference.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;amountInCents&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;typeof&lt;/span&gt; &lt;span class="n"&gt;value&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s2"&gt;"number"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;value&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;typeof&lt;/span&gt; &lt;span class="n"&gt;value&lt;/span&gt; &lt;span class="o"&gt;!==&lt;/span&gt; &lt;span class="s2"&gt;"string"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;money&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt;
    &lt;span class="o"&gt;/^&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;\&lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="o"&gt;+&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;?:&lt;/span&gt;\&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;\&lt;span class="n"&gt;d&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="o"&gt;?&lt;/span&gt;$&lt;span class="o"&gt;/&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;match&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;money&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;exec&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="n"&gt;match&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt;

  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;fraction&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;match&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="s2"&gt;""&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;digits&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt;
    &lt;span class="n"&gt;match&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;fraction&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;padEnd&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"0"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;cents&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Number&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;digits&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;Number&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isSafeInteger&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cents&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;cents&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="nb"&gt;console&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;assert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="nf"&gt;amountInCents&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"2.75"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="mi"&gt;275&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="nb"&gt;console&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;assert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="nf"&gt;amountInCents&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;2.75&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="mi"&gt;275&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="nb"&gt;console&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;assert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="nf"&gt;amountInCents&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"2.751"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;undefined&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  A second check on the original order
&lt;/h2&gt;

&lt;p&gt;We linked the older journal entry to the order using the original ticket link and verified that the order’s ticket reference matched the PDF sent to Telegram. ARGUS fetched the order JSON and receipt PDF from Coach USA with two GET requests. The receipt said &lt;strong&gt;Captured&lt;/strong&gt;, with &lt;strong&gt;$2.75 paid&lt;/strong&gt; and the expected card suffix.&lt;/p&gt;

&lt;p&gt;On the issuer side, Laso showed &lt;strong&gt;one new matching debit for $2.75&lt;/strong&gt;. Available funds moved from &lt;strong&gt;$5.00 to $2.25&lt;/strong&gt;. ARGUS saved the order number, a SHA-256 digest of the receipt, and the reconciliation result. A repeated check returned the saved result for the same purchase. No second order was sent. Laso’s card activity did not expose final settlement status, so the evidence here is a captured merchant receipt and a matching issuer debit. &lt;a href="https://help.laso.finance/en/articles/12421793-i-was-waiting-on-a-refund-on-a-transaction-authorization-now-it-s-gone-what-happened" rel="noopener noreferrer"&gt;How Laso reports card authorizations&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;For a new purchase, ARGUS records the payment attempt before it submits the order and stores the merchant order number as soon as it has it. If the browser disconnects around payment, the purchase stays open for reconciliation. A fresh checkout would risk a second charge. A lookup can use the order already on file.&lt;/p&gt;

&lt;p&gt;This pilot covered one route, one passenger and one card. That’s enough to test the boundary I care about in any agent checkout: an uncertain payment needs a place to wait and a way to look up what happened.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/examples?from=devto-hudson#first-call" rel="noopener noreferrer"&gt;Get a free eSIM quote for the Netherlands&lt;/a&gt;. Run one command to get up to two plans and a signed Hub receipt. You can run it without an API key or wallet.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;I used AI assistance to draft and edit this write-up.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>payments</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Selling eSIMs to AI agents: the seller's side, tried on a laptop in 32 seconds</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Fri, 13 Nov 2026 21:27:31 +0000</pubDate>
      <link>https://dev.to/alexar76/selling-esims-to-ai-agents-the-sellers-side-tried-on-a-laptop-in-32-seconds-560d</link>
      <guid>https://dev.to/alexar76/selling-esims-to-ai-agents-the-sellers-side-tried-on-a-laptop-in-32-seconds-560d</guid>
      <description>&lt;p&gt;If you resell travel eSIMs, you probably have a wholesale account with eSIM Access, eSIM Go or Airalo Partners, and a checkout built for people: pick a country, pay by card, scan a QR code. An AI agent that books a trip cannot use that checkout. It needs an endpoint that answers "3 GB in Japan" with priced offers, takes a payment without an account, and returns the activation code in the same response.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://github.com/alexar76/esim-operator-kit" rel="noopener noreferrer"&gt;eSIM operator kit&lt;/a&gt; (MIT) is that endpoint plus a storefront for people, built around your existing supplier account. Until this week, though, it asked for a lot before showing anything. The first step wanted a server, a domain with an A record, free ports 80 and 443, and your supplier's API keys. We cloned our own repository as a stranger would, ran &lt;code&gt;./onboard.sh&lt;/code&gt; on a laptop, and it stopped at &lt;code&gt;.env is missing&lt;/code&gt; and asked for a domain. A reseller who only wanted to see what agents would actually call had to register with a supplier first.&lt;/p&gt;

&lt;p&gt;So the first step is now a command that needs none of that.&lt;/p&gt;

&lt;h2&gt;
  
  
  One command, no accounts
&lt;/h2&gt;

&lt;p&gt;You need git and Docker. We ran it with Docker Engine 29.2.1 and Compose 5.1.0; the overlay for this mode uses Compose's &lt;code&gt;!reset&lt;/code&gt; and &lt;code&gt;!override&lt;/code&gt; merge tags, which old Compose v2 releases do not understand.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/alexar76/esim-operator-kit.git
&lt;span class="nb"&gt;cd&lt;/span&gt; esim-operator-kit
./onboard.sh --try
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It builds two containers and binds them to &lt;code&gt;127.0.0.1&lt;/code&gt; only: the storefront on port 9495 and the seller node, the part agents talk to, on port 9490. The supplier behind them is &lt;code&gt;SUPPLIER=sample&lt;/code&gt;, a catalogue of 49 made-up plans across 26 countries with invented costs. The kit refuses to start that supplier unless &lt;code&gt;DEMO_MODE=1&lt;/code&gt; is set, and the storefront banner says the plans and prices are made up. Nothing is written to &lt;code&gt;.env&lt;/code&gt;, &lt;code&gt;secrets/&lt;/code&gt; or &lt;code&gt;data/&lt;/code&gt;: the containers keep their state on tmpfs.&lt;/p&gt;

&lt;p&gt;On an Apple M4 with 32 GB this took 32 seconds with a cold Docker build and 6 seconds once the images were cached. At the end the script runs the same conformance check that any seller node has to pass before it joins the federation. This is its output from that run, unedited:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;✓ listing.json — esim.quote@v1; order at http://localhost:9490/order
✓ esim.quote@v1 answer — 3 offers for JP, all valid
✓ cheapest first
✓ signature — Ed25519 over the request-bound canonical, key from listing.json
✓ bad input is a 400
✓ unknown eSIM is a 404 not_found
✓ unknown offer is a 400 before any price
⚠ x402 terms — a demo node: it sells nothing (demo_mode)
✓ CORS on the free routes

Conforms to esim.*@v1.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The warning is expected. A node in demo mode has no payment terms to check, because it never sells.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;./onboard.sh --try --down&lt;/code&gt; removes the containers and the images it built. Docker's build cache is all that is left.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/selling-esims-to-ai-agents-tried-on-a-laptop-in-32-seconds/what-runs-where.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fhnzvo2wgtkyskkczn7ki.png" alt="Two stacked panels. Top, your laptop running ./onboard.sh --try: a storefront on localhost:9495 and a seller node on localhost:9490, both reading SUPPLIER=sample with 49 made-up plans, and esim-rails check with 8 checks passing and one warning for the demo node; DEMO_MODE=1, state in memory, nothing written, no payments, no hub announcement; 32 seconds cold, 6 warm. Below, your server running ./onboard.sh: Caddy with TLS in front of the storefront and the seller node, a federation node listed on AIMarket hubs, your supplier, and x402 settling USDC on Base into your wallet." width="800" height="1225"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The same code on both sides. &lt;code&gt;--try&lt;/code&gt; swaps the supplier for a sample catalogue and leaves out TLS, payments and the hub announcement.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What an agent sends
&lt;/h2&gt;

&lt;p&gt;The quote is a plain POST, free and unauthenticated:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl -s localhost:9490/quote -H content-type:application/json \
  -d &lt;span class="s1"&gt;'{"country":"JP","min_gb":3}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This is the response from the sample catalogue, shortened. The second offer and some fields are cut:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"success"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"result"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"seller"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Sample eSIM store"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"product_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"esim-sample"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"demo"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"offers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
      &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"offer_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"SAMPLE-JP-3-15"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Sample · Japan 3 GB / 15 days"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"coverage"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"kind"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"country"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"countries"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"JP"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"data_gb"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"days"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;15&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"price"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"usd"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"6.00"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"currency"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"USDC"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"usd_per_gb"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"order"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"http://localhost:9490/order"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"method"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"POST"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"payment"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"x402, USDC on Base"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Offers come cheapest first, and the node signs each answer with its Ed25519 key, which it publishes in its &lt;code&gt;listing.json&lt;/code&gt;. An agent that found the seller through a hub can check that the quote came from the node the hub listed.&lt;/p&gt;

&lt;p&gt;Then the agent orders:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl -si localhost:9490/order -X POST -H content-type:application/json \
  -d &lt;span class="s1"&gt;'{"offer_id":"SAMPLE-JP-3-15"}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In &lt;code&gt;--try&lt;/code&gt; it ends here, on purpose:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;HTTP/1.1 403 Forbidden

{"success":false,"error":"demo_mode","message":"demo mode: this node quotes made-up sample prices and sells nothing; connect a wholesale account to sell"}
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Until 13 November that message said "quotes live prices", which was true on our public demo node and false on a laptop running sample plans. Writing this article is how we noticed. The node now says which of the two it is.&lt;/p&gt;

&lt;h2&gt;
  
  
  The same order on a node that sells
&lt;/h2&gt;

&lt;p&gt;With a real supplier and demo mode off, that first &lt;code&gt;POST /order&lt;/code&gt; returns &lt;code&gt;402 Payment Required&lt;/code&gt; with &lt;a href="https://www.x402.org/" rel="noopener noreferrer"&gt;x402&lt;/a&gt; terms: the amount, your wallet address, USDC on Base. The agent signs one EIP-3009 transfer authorization for exactly that amount and sends the order again with a &lt;code&gt;PAYMENT-SIGNATURE&lt;/code&gt; header. It needs no account with you, no API key, and no ETH for gas.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/selling-esims-to-ai-agents-tried-on-a-laptop-in-32-seconds/agent-purchase.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fsl7s18wljf5bb6kov07n.png" alt="Sequence diagram with five lanes: AI agent, hub, your seller node, x402 facilitator, supplier. The agent searches the hub for intent=esim and gets the node's listing; posts a quote for JP, 3 GB and gets signed offers, cheapest first; posts an order and gets 402 Payment Required with amount, payTo and USDC on Base; signs one EIP-3009 authorization without gas and posts the order again with PAYMENT-SIGNATURE. The node asks the facilitator to verify, orders from the supplier with a transaction id derived from the payment, so the same payment twice gives the same eSIM, receives the activation code and ICCID, answers 200 with the activation code, a status token and a signed receipt, then settles USDC to your wallet. Two notes: if the supplier refuses, the answer is 409 and nothing is settled; in --try and on any demo node the first order stops at 403 demo_mode." width="799" height="631"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;An agent's purchase from a selling node, as &lt;code&gt;rails/src/node.js&lt;/code&gt; handles it. The money moves last.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Three things in that path matter to an operator more than to the agent:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The money moves after the eSIM exists.&lt;/strong&gt; The node asks the facilitator to verify the signature, orders the profile from your supplier, answers with the activation code, and only then settles. If your supplier refuses the order, the agent gets &lt;code&gt;409 "the supplier refused this order; you were not charged"&lt;/code&gt;, and an error answer is never settled.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A retry cannot buy twice.&lt;/strong&gt; The supplier's transaction id is derived from the payment itself. An agent that times out and sends the same signed payment again gets the same eSIM, not a second one billed to your wholesale balance.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The activation code goes to the payer only.&lt;/strong&gt; It is in the order response and nowhere else: not in webhooks, logs or receipts. The status token lets the buyer check installation or ask for a refund of an unused eSIM later.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This path has run with real money. In October our own agent bought a Turkey eSIM through it for 0.31 USDC and got the money back for the unused profile. Both transfers are on Base, and &lt;a href="https://modelmarket.dev/blog/our-ai-agent-bought-an-esim-for-its-owner-then-got-the-money-back-every-step-is-on-chain/" rel="noopener noreferrer"&gt;that article&lt;/a&gt; has the transaction links.&lt;/p&gt;

&lt;h2&gt;
  
  
  From the laptop to a real account
&lt;/h2&gt;

&lt;p&gt;On a server, the same script without &lt;code&gt;--try&lt;/code&gt; asks for the domain, the store name and your supplier keys. It writes &lt;code&gt;.env&lt;/code&gt;, starts in demo mode, and announces the node to a hub, so you can check everything before anything is sold.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Supplier&lt;/th&gt;
&lt;th&gt;&lt;code&gt;SUPPLIER=&lt;/code&gt;&lt;/th&gt;
&lt;th&gt;Status&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;eSIM Access&lt;/td&gt;
&lt;td&gt;&lt;code&gt;esimaccess&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;live-tested: real orders, cancellations and refunds&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Airalo Partner API&lt;/td&gt;
&lt;td&gt;&lt;code&gt;airalo&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;built to Airalo's published API reference; not yet run on a live Airalo account&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;eSIM Go API v2.5&lt;/td&gt;
&lt;td&gt;&lt;code&gt;esimgo&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;built to eSIM Go's published reference; not yet run on a live account&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Your own platform&lt;/td&gt;
&lt;td&gt;&lt;code&gt;http&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;five signed HTTP endpoints: catalog, orders, profiles, cancel, balance&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;If you already run Airalo or eSIM Go, you would be the first to run those adapters against a live account, and we would rather hear what breaks than guess.&lt;/p&gt;

&lt;p&gt;By default the price is your supplier's cost plus 100%, and never less than cost plus $1.00. A price list (&lt;code&gt;PRICES_FILE&lt;/code&gt;, CSV or JSON) overrides that per offer, country, region or plan size. You remain the merchant of record: VAT, refunds policy and support are yours, as they are on your current storefront. Agents pay in USDC on Base. The storefront for people takes crypto on ten networks, and card payments can be added with your own Stripe keys.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this stands
&lt;/h2&gt;

&lt;p&gt;We run one demo seller node on &lt;a href="https://modelmarket.dev/providers/esim/healthz" rel="noopener noreferrer"&gt;modelmarket.dev&lt;/a&gt;. Its catalogue comes live from eSIM Access, 1,823 plans in 205 countries when we checked on 13 November, and it refuses every order. No eSIM operator outside our team has connected a node to the federation yet, and the only agent purchases so far are our own tests.&lt;/p&gt;

&lt;p&gt;That is also why the cheap first step matters. Before this week, finding out whether the agent side fits your business meant opening a supplier account and renting a server. Now it costs you 32 seconds on a laptop.&lt;/p&gt;

&lt;p&gt;The kit, the guides in five languages and the conformance check are in the &lt;a href="https://github.com/alexar76/esim-operator-kit" rel="noopener noreferrer"&gt;repository&lt;/a&gt;. The landing at &lt;a href="https://esim.modelmarket.dev/" rel="noopener noreferrer"&gt;esim.modelmarket.dev&lt;/a&gt; has the same three commands at the top.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We build the eSIM operator kit and run the modelmarket.dev hub. AI assistance was used for the code changes, the measurements, the figures and this article. The commands and outputs above come from runs on 13 November 2026 and can be repeated from the repository.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>docker</category>
      <category>payments</category>
    </item>
    <item>
      <title>Six of 16 common MCP servers can't be read from outside. One of the others changed 19 times in a month</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Thu, 12 Nov 2026 06:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/six-of-16-common-mcp-servers-cant-be-read-from-outside-one-of-the-others-changed-19-times-in-a-1dh4</link>
      <guid>https://dev.to/alexar76/six-of-16-common-mcp-servers-cant-be-read-from-outside-one-of-the-others-changed-19-times-in-a-1dh4</guid>
      <description>&lt;p&gt;We wrote down sixteen MCP servers that a software team could plausibly connect to its agents: the hosted servers of&lt;br&gt;
GitHub, Notion, Linear, Atlassian, Sentry and Stripe; Context7, DeepWiki, Hugging Face and Cloudflare's docs server&lt;br&gt;
for reading; Firecrawl, Brave Search and Tavily for the web; and three local packages, Playwright, MongoDB and Sentry.&lt;br&gt;
It is a sample we picked, not a customer's setup and not a survey.&lt;/p&gt;

&lt;p&gt;Then we asked &lt;a href="https://histor.modelmarket.dev" rel="noopener noreferrer"&gt;HISTOR&lt;/a&gt;, the public log of MCP tool definitions we run, what each of&lt;br&gt;
them had told agents over the last 30 days, 12 October to 11 November 2026.&lt;/p&gt;

&lt;p&gt;For six of the sixteen the answer was: nothing. HISTOR has never read a tool from them.&lt;/p&gt;
&lt;h2&gt;
  
  
  The six nobody outside can read
&lt;/h2&gt;

&lt;p&gt;GitHub, Notion, Atlassian and Stripe answer an unauthenticated client with HTTP 401. Linear's and Sentry's hosted&lt;br&gt;
servers did the same on every read until 1 November. Since 2 November HISTOR honours robots.txt, and both sites&lt;br&gt;
disallow &lt;code&gt;/mcp&lt;/code&gt; for all crawlers (Linear's disallows the whole site), so HISTOR no longer sends them a request.&lt;/p&gt;

&lt;p&gt;Requiring sign-in is the right call for servers that reach a team's code, tickets, documents and payments. The&lt;br&gt;
consequence is that no public log, ours included, can tell you what these servers describe to your agent, or when&lt;br&gt;
that description changed. The readable servers are not harmless by comparison — MongoDB's package has&lt;br&gt;
&lt;code&gt;drop-database&lt;/code&gt;, Sentry's package now has &lt;code&gt;delete_alert_rule&lt;/code&gt; — but at least their definitions are public.&lt;/p&gt;
&lt;h2&gt;
  
  
  The ten that can be read
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/six-of-16-common-mcp-servers-cant-be-read-from-outside/changes-by-server.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fmodelmarket.dev%2Fblog%2Fsix-of-16-common-mcp-servers-cant-be-read-from-outside%2Fchanges-by-server.png" alt="Bar chart, a sample stack of 16 MCP servers, changes in 30 days: Firecrawl 19, of which 13 at the same version; Playwright npm and Hugging Face 2; Sentry npm, MongoDB npm, Cloudflare docs, DeepWiki and Context7 1; Tavily npm and Brave Search npm none. GitHub, Notion, Linear, Atlassian, Sentry hosted and Stripe are not readable from outside." width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Eight of the ten changed their tool definitions — descriptions or schemas — 28 times between them. Brave Search and&lt;br&gt;
Tavily did not change at all. Thirteen of the 28 changes came while the server kept reporting the same version, and&lt;br&gt;
all thirteen were Firecrawl's. Everywhere else, a change came with a new version number.&lt;/p&gt;

&lt;p&gt;A changed definition does not mean a server misbehaved. It means the text and the schemas your model works from are&lt;br&gt;
not the ones somebody reviewed, and the workflows that use those tools deserve a re-run. Three of the ten show three&lt;br&gt;
different kinds of change.&lt;/p&gt;
&lt;h3&gt;
  
  
  Sentry: from 9 tools to 73
&lt;/h3&gt;

&lt;p&gt;Between the two versions HISTOR read, &lt;code&gt;@sentry/mcp-server&lt;/code&gt; 0.42.0 and 0.49.0, the package went from 9 tools to 73.&lt;br&gt;
It dropped &lt;code&gt;execute_sentry_tool&lt;/code&gt;, &lt;code&gt;search_sentry_tools&lt;/code&gt; and &lt;code&gt;search_events&lt;/code&gt;, kept six, and added 67. Among the new&lt;br&gt;
ones: &lt;code&gt;delete_alert_rule&lt;/code&gt;, &lt;code&gt;delete_metric_monitor&lt;/code&gt;, &lt;code&gt;delete_monitor_environment&lt;/code&gt;, &lt;code&gt;delete_uptime_monitor&lt;/code&gt;, seven&lt;br&gt;
&lt;code&gt;create_*&lt;/code&gt; and five &lt;code&gt;update_*&lt;/code&gt; tools.&lt;/p&gt;

&lt;p&gt;The old design was deliberate. In 0.42.0, &lt;code&gt;search_sentry_tools&lt;/code&gt; told the model: "Many Sentry operations are&lt;br&gt;
intentionally not exposed as top-level tools", and &lt;code&gt;execute_sentry_tool&lt;/code&gt; ran them by name. In 0.49.0 those operations&lt;br&gt;
are tools of their own. For a reviewer that is better: the four &lt;code&gt;delete_*&lt;/code&gt; tools are now visible by name. But a&lt;br&gt;
config that starts &lt;code&gt;npx -y @sentry/mcp-server&lt;/code&gt; without a version moved to 0.49.0 as soon as npx resolved it, and&lt;br&gt;
nobody had to approve the new tool list.&lt;/p&gt;
&lt;h3&gt;
  
  
  Firecrawl: changes the version did not announce
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/six-of-16-common-mcp-servers-cant-be-read-from-outside/firecrawl-timeline.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fmodelmarket.dev%2Fblog%2Fsix-of-16-common-mcp-servers-cant-be-read-from-outside%2Ffirecrawl-timeline.png" alt="Timeline of Firecrawl's 19 changes between 12 October and 11 November 2026. Six came with a new version: 3.30.0, 3.30.1, 3.30.3, 3.34.0, 3.38.0 and 3.39.0. Thirteen came at an unchanged version, four of them on 10 and 11 November." width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Each line is a new tool set: HISTOR hashes the tools' names, descriptions and input and output schemas, and a new&lt;br&gt;
digest is a change. The orange lines came with no new version. Some of what they carried:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;an &lt;code&gt;objective&lt;/code&gt; input became required on 13 October, stopped being required with the 3.30.1 release on 15
October, and became required again on 18 October while the server still said 3.30.3;&lt;/li&gt;
&lt;li&gt;on 10 November &lt;code&gt;firecrawl_scrape&lt;/code&gt; gained an output field, &lt;code&gt;retryable&lt;/code&gt;, and the &lt;code&gt;zeroDataRetention&lt;/code&gt; option
disappeared from &lt;code&gt;firecrawl_search&lt;/code&gt;'s scrape options — an object that rejects unknown keys, so a call that still
sends the flag no longer validates;&lt;/li&gt;
&lt;li&gt;at 03:21 UTC on 11 November &lt;code&gt;firecrawl_scrape&lt;/code&gt; dropped its &lt;code&gt;alexandria&lt;/code&gt; argument, made &lt;code&gt;url&lt;/code&gt; required, and its
description named &lt;code&gt;alexandria_library_search&lt;/code&gt; and &lt;code&gt;alexandria_library_read&lt;/code&gt; where it had named
&lt;code&gt;firecrawl_find_tools&lt;/code&gt;. At 07:58 it was back, byte for byte, to the definition it had before; &lt;code&gt;firecrawl_search&lt;/code&gt;
kept the new names. The server reported 3.39.0 throughout.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We don't know why any of these happened, and nothing about them is wrong. Firecrawl is under active development and&lt;br&gt;
its hosted server is where that shows. What they show is that the version string of a hosted server is not a record&lt;br&gt;
of what it serves. (Yesterday's post on &lt;a href="https://modelmarket.dev/blog/same-mcp-version-different-tool-description-who-signed-it/" rel="noopener noreferrer"&gt;signed manifest&lt;br&gt;
revisions&lt;/a&gt; is about who&lt;br&gt;
can vouch for such a change.)&lt;/p&gt;
&lt;h3&gt;
  
  
  MongoDB: a change that changes nothing
&lt;/h3&gt;

&lt;p&gt;&lt;code&gt;mongodb-mcp-server&lt;/code&gt; 3.2.2 → 3.3.0 touched 10 input schemas and 8 output schemas. Most of its 47 edits rewrite&lt;br&gt;
&lt;code&gt;additionalProperties: {}&lt;/code&gt; as &lt;code&gt;additionalProperties: true&lt;/code&gt; or drop an empty &lt;code&gt;items: {}&lt;/code&gt;. JSON Schema treats these&lt;br&gt;
forms as identical, so every call that validated before still validates. The rest add guidance on embedding models to&lt;br&gt;
field descriptions and tighten one field mapping in &lt;code&gt;create-index&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;A digest-based alarm fires on this exactly as loudly as on Firecrawl's &lt;code&gt;zeroDataRetention&lt;/code&gt;. That is why a change&lt;br&gt;
alert alone is not enough: the person approving it needs the diff.&lt;/p&gt;
&lt;h2&gt;
  
  
  Checking your own servers
&lt;/h2&gt;

&lt;p&gt;HISTOR's API is public and needs no key. This is the count behind the orange lines, cut down to one server; it runs on&lt;br&gt;
Python 3.9 or later with the standard library:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# usage: python3 same_version.py &amp;lt;server id&amp;gt; &amp;lt;since, e.g. 2026-10-12&amp;gt;&lt;/span&gt;
&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;sys&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;urllib&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;

&lt;span class="n"&gt;server_id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;since&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;sys&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;argv&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;sys&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;argv&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="n"&gt;api&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"https://histor.modelmarket.dev/api/v1"&lt;/span&gt;
&lt;span class="n"&gt;url&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="s2"&gt;"{api}/changes?watch={server_id}&amp;amp;limit=200"&lt;/span&gt;  &lt;span class="c1"&gt;# newest first&lt;/span&gt;
&lt;span class="n"&gt;rows&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;load&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;urllib&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;urlopen&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;url&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;30&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"changes"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

&lt;span class="n"&gt;window&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;rows&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"observed_at"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="n"&gt;since&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="n"&gt;silent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;window&lt;/span&gt;
          &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"from_server_version"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"to_server_version"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
          &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"from_package_version"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"to_package_version"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

&lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="s2"&gt;"{len(window)} changes since {since}, {len(silent)} at the same version"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;silent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;tools&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"tool"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;m&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"summary"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"modified"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;" "&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"observed_at"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"to_server_version"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;", "&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ python3 same_version.py 0ae1ed4e4d5d8dd7 2026-10-12
19 changes since 2026-10-12, 13 at the same version
  2026-11-11T07:58:19Z 3.39.0 firecrawl_scrape
  2026-11-11T03:21:09Z 3.39.0 firecrawl_scrape, firecrawl_search
  …
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A server's id is on its HISTOR page, or from &lt;code&gt;https://histor.modelmarket.dev/api/v1/servers?q=&amp;lt;name&amp;gt;&lt;/code&gt;. The numbers&lt;br&gt;
move as servers change; the output above is from 11 November.&lt;/p&gt;

&lt;p&gt;For what HISTOR cannot read, the check has to run with your credentials. &lt;a href="https://www.npmjs.com/package/@aimarket/warden" rel="noopener noreferrer"&gt;WARDEN&lt;/a&gt;,&lt;br&gt;
our open-source MCP scanner, records the tool definitions of every server in a repository's &lt;code&gt;.mcp.json&lt;/code&gt; into a lock&lt;br&gt;
file. It needs Node.js 20 or later, and it starts the local (stdio) servers in the config to read their tools, so run&lt;br&gt;
it where you would run those servers anyway:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx -y @aimarket/warden@0.26.0 scan --project \
  --lock warden.lock.json --update-lock
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;WARDEN scan 0.26.0 · ruleset 19 sha256-XfRh1+VZoAM6… · block at high
  read .mcp.json (claude-code, 2 servers)

  ✓ allow   deepwiki  claude-code    3 tools · score 0.90 · lock: updated
  ✓ allow   context7  claude-code    2 tools · score 0.90 · lock: updated
…
Lock written: warden.lock.json. Commit it; the diff is the review.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That run is from a test repository with two public servers, on 11 November. Committed, the lock makes CI fail when a&lt;br&gt;
remote server serves different definitions (&lt;code&gt;TOOL_DEF_DRIFT&lt;/code&gt;, with a field-by-field diff) or when a server nobody&lt;br&gt;
approved appears (&lt;code&gt;LOCK_MISSING&lt;/code&gt;); &lt;code&gt;--project&lt;/code&gt; keeps your personal configs out of it.&lt;/p&gt;

&lt;p&gt;It reaches a hosted server only with the headers written in the config: a server that takes a token in a header&lt;br&gt;
(GitHub's accepts a personal access token, for example) can be locked this way. A server you sign into through OAuth&lt;br&gt;
in the client answers WARDEN with 401, just as it answers HISTOR. For those, today, the review is the tool list your&lt;br&gt;
client shows you. We don't know of a tool that closes that gap, ours included.&lt;/p&gt;

&lt;p&gt;For package servers the oldest fix still works: write the version you reviewed into the config,&lt;br&gt;
&lt;code&gt;"args": ["-y", "@sentry/mcp-server@0.49.0"]&lt;/code&gt;, and review again when you bump it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits of this measurement
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;HISTOR reads remote servers on a six-hour target, and every 30 minutes once a server has changed twice in a week.
At 14:44 UTC on 11 November the mean gap between reads was 13.7 hours. A change that is undone between two reads
is not seen.&lt;/li&gt;
&lt;li&gt;It reads what a server serves an unauthenticated client. Tools that appear only after sign-in, or only for
another client or launch configuration, are outside what it can see.&lt;/li&gt;
&lt;li&gt;Packages are read at whatever version is current when a crawl reaches them, every few days for the ones here.
HISTOR never read Sentry 0.43–0.47, and its 0.48.0 install failed, so their changes are inside the 0.42 → 0.49
difference.&lt;/li&gt;
&lt;li&gt;Sixteen servers is a sample we chose. At 14:44 UTC on 11 November HISTOR counted 79,415 tool-set changes in the
previous 30 days across the 91,847 registry entries it tracks; that total includes publishers whose servers change
on almost every read, so it is not a rate for your stack.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The sixteen records as we read them at 14:20 UTC on 11 November, their SHA-256 sums, the stats snapshot and the&lt;br&gt;
scripts behind every number here are in &lt;a href="https://modelmarket.dev/blog/six-of-16-common-mcp-servers-cant-be-read-from-outside/stack-audit-data-2026-11-11.zip" rel="noopener noreferrer"&gt;one archive&lt;/a&gt;&lt;br&gt;
(1 MB).&lt;/p&gt;

&lt;h2&gt;
  
  
  If you'd rather someone else did this
&lt;/h2&gt;

&lt;p&gt;We are starting to do this as an audit for teams: an inventory of the MCP servers your developers and agents use,&lt;br&gt;
a WARDEN scan of each tool, the lock and an allowlist in your CI, and 30 days of change alerts with a short note on&lt;br&gt;
each change that touches your workflows. The first three teams get it free in exchange for written feedback and a&lt;br&gt;
short, non-binding letter of intent. If that's you, write to &lt;a href="mailto:qwezzy721@gmail.com"&gt;qwezzy721@gmail.com&lt;/a&gt; with the MCP clients your team uses.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We run HISTOR and WARDEN. AI assistance was used for the analysis, the scripts, the figures and this article; the&lt;br&gt;
data, code and quoted outputs are reproducible from the linked files.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>security</category>
      <category>ai</category>
      <category>devops</category>
    </item>
    <item>
      <title>Same MCP version. Different tool description. Who signed it?</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Wed, 11 Nov 2026 07:27:51 +0000</pubDate>
      <link>https://dev.to/alexar76/same-mcp-version-different-tool-description-who-signed-it-5fbd</link>
      <guid>https://dev.to/alexar76/same-mcp-version-different-tool-description-who-signed-it-5fbd</guid>
      <description>&lt;p&gt;On Monday, a test MCP server calls its tool &lt;code&gt;search&lt;/code&gt; and describes it as “Search the docs.” On Tuesday, it says “Search the private docs.” The server still reports version &lt;code&gt;1.4.2&lt;/code&gt;. The tool name and input schema are identical. Would your client ask for approval again?&lt;/p&gt;

&lt;p&gt;An MCP client discovers tool descriptions and argument schemas through &lt;code&gt;tools/list&lt;/code&gt;, and a host can pass that metadata to its model. Whoever controls those strings can influence what the model thinks a tool does before a call ever happens. Invariant Labs demonstrated this with tool poisoning. A November 2026 preprint by Farooq Shaikh and Biagio Montaruli tested poisoning across agent frameworks and checked whether attacks led to actual tool execution. &lt;a href="https://modelcontextprotocol.io/specification/2026-07-28/server/tools#listing-tools" rel="noopener noreferrer"&gt;MCP tools specification&lt;/a&gt; · &lt;a href="https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks" rel="noopener noreferrer"&gt;Invariant Labs disclosure&lt;/a&gt; · &lt;a href="https://arxiv.org/abs/2611.03555v2" rel="noopener noreferrer"&gt;Cross-framework study, v2&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;HISTOR already had a way to record what it observed a server advertising. The new code handles the publisher’s side of the story. It lets a release workflow sign the tool manifest and gives a client a way to compare the signed revision with the definitions it just received. The format follows a proposed test in the OWASP MCP Security Testing Guide. &lt;a href="https://github.com/OWASP/MCP-Security-Testing-Guide/pull/21" rel="noopener noreferrer"&gt;Draft MCPSTG-SUPP-06&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  One sentence, two new hashes
&lt;/h2&gt;

&lt;p&gt;The publisher takes a snapshot of &lt;code&gt;tools/list&lt;/code&gt;, following the pagination cursor until the list is complete. The seven fields eligible for HISTOR’s revision format are &lt;code&gt;name&lt;/code&gt;, &lt;code&gt;title&lt;/code&gt;, &lt;code&gt;description&lt;/code&gt;, &lt;code&gt;inputSchema&lt;/code&gt;, &lt;code&gt;outputSchema&lt;/code&gt;, &lt;code&gt;annotations&lt;/code&gt;, and &lt;code&gt;execution&lt;/code&gt;. It uses tool names as object keys and rejects duplicates so one entry can’t silently overwrite another.&lt;/p&gt;

&lt;p&gt;Each entry gets a SHA-256 digest. The full name-keyed object gets one too. Both hash the bytes of canonical JSON, so a reordered tool list or JSON object produces the same digest. The package version travels in the signed statement as context. The manifest digest depends on the tool definitions. &lt;a href="https://www.rfc-editor.org/rfc/rfc8785" rel="noopener noreferrer"&gt;RFC 8785&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/m8608wpp7yda36b76ixd.png" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm8608wpp7yda36b76ixd.png" alt="Two captures show search at version 1.4.2. The description changes from “Search the docs.” to “Search the private docs.” The per-tool and manifest hashes both change." width="800" height="817"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;One synthetic tool, one description edit. The diagram shows the first eight hexadecimal characters of each SHA-256 digest.&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Run the one-line edit
&lt;/h3&gt;

&lt;p&gt;Use the pinned HISTOR source checkout with its project dependencies installed. From that checkout, save the block as &lt;code&gt;compare.py&lt;/code&gt; and run &lt;code&gt;python compare.py&lt;/code&gt; in your project environment. The example was checked with Python 3.12.13 and uses two synthetic tool lists. &lt;a href="https://github.com/alexar76/histor/commit/ba957e40a489440b3f85f33f1a072254a239624b" rel="noopener noreferrer"&gt;HISTOR source at ba957e4&lt;/a&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;histor&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;revision&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;build_statement&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;digests&lt;/span&gt;

&lt;span class="n"&gt;tool&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="s2"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"search"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Search the docs."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"inputSchema"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="s2"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"object"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"properties"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"query"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"string"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"required"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"query"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="n"&gt;changed&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;tool&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Search the private docs."&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="n"&gt;release&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="s2"&gt;"server"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"io.github.example/docs"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"identity"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="s2"&gt;"https://github.com/example/docs/"&lt;/span&gt;
        &lt;span class="s2"&gt;".github/workflows/sign.yml@refs/heads/main"&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"server_version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"1.4.2"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="n"&gt;before&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;build_statement&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;tool&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;release&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;after&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;build_statement&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;changed&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;previous&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;before&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;release&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="n"&gt;before&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"serverVersion"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;after&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"serverVersion"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="s2"&gt;"1.4.2"&lt;/span&gt;
&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="n"&gt;after&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"previousManifestDigest"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;before&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"manifestDigest"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="n"&gt;before&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"manifestDigest"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="n"&gt;after&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"manifestDigest"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="n"&gt;before&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"toolDigests"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"search"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="n"&gt;after&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"toolDigests"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"search"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;snapshot&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;tool&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;changed&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;manifest&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;per_tool&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;digests&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;snapshot&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;per_tool&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"search"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;7&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;15&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;manifest&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;7&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;15&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first column is the tool digest prefix. The second is the manifest digest prefix.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;896aa19c 4f97f46c
1d051183 cdf20eeb
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  What the workflow actually signs
&lt;/h2&gt;

&lt;p&gt;CI signs a small statement containing the server name, revision number, manifest digest, per-tool digests, and a link to the previous manifest digest. &lt;code&gt;histor revision build&lt;/code&gt; writes the statement as canonical bytes. The verifier reconstructs those same bytes from the envelope by removing &lt;code&gt;signer.signature&lt;/code&gt; and &lt;code&gt;transparencyLog&lt;/code&gt;. The envelope can be formatted for humans because the verifier canonicalizes the statement again.&lt;/p&gt;

&lt;p&gt;The publisher’s GitHub Actions workflow uses OIDC to obtain a short-lived signing certificate from Sigstore’s Fulcio. For an &lt;code&gt;io.github.&amp;lt;owner&amp;gt;/...&lt;/code&gt; name, the HISTOR verifier checks the GitHub Actions issuer and the certificate’s repository owner against that namespace. The CLI can also pin an exact workflow identity. Rekor’s inclusion proof is a separate check on the signed entry. HISTOR verifies the proof from the Sigstore bundle and checks that its log ID and index agree with the copy in the envelope. &lt;a href="https://docs.sigstore.dev/cosign/signing/overview/" rel="noopener noreferrer"&gt;Sigstore signing overview&lt;/a&gt; · &lt;a href="https://docs.sigstore.dev/logging/overview/" rel="noopener noreferrer"&gt;Rekor&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The difference shows up when a signed revision is valid but a client receives different tools. Signature verification still succeeds. For a listed server with an ingested revision, a valid raw &lt;code&gt;tools&lt;/code&gt; array in &lt;code&gt;POST /api/v1/check&lt;/code&gt; lets HISTOR recompute the digest. In this case, &lt;code&gt;signedRevision.match&lt;/code&gt; comes back as &lt;code&gt;different&lt;/code&gt;. When the client sends only an MTL/1 digest, HISTOR omits that comparison. MTL/1 labels describe observer snapshots and use their own digest format.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/nymzw5xqgf4s2lo2agua.png" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fnymzw5xqgf4s2lo2agua.png" alt="Publisher CI signs a revision using Sigstore. HISTOR verifies the bundle and Rekor proof at ingest, records the accepted revision, and compares its manifest digest with the client’s raw tools in a separate step." width="800" height="1111"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Rekor records the publisher’s signature. HISTOR’s revision tree records its own acceptance of that revision. The client’s tool list is compared when it calls &lt;code&gt;/check&lt;/code&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What the pilot covered
&lt;/h2&gt;

&lt;p&gt;On November 11, the first publisher signature covered revision &lt;code&gt;1&lt;/code&gt; of &lt;code&gt;io.github.alexar76/histor&lt;/code&gt; at Rekor index &lt;code&gt;3803861704&lt;/code&gt;. To reproduce it, the local verifier checked the public bundle against the expected workflow identity and compared the signed digest with a fresh listing of the three tools from the released &lt;code&gt;aimarket-histor&lt;/code&gt; 0.7.0 package. Both checks passed. Editing the version in the envelope broke signature verification. &lt;a href="https://raw.githubusercontent.com/alexar76/histor/manifest-revisions/revisions/1.json" rel="noopener noreferrer"&gt;Signed revision&lt;/a&gt; · &lt;a href="https://github.com/alexar76/histor/actions/runs/42968858469" rel="noopener noreferrer"&gt;Signing run&lt;/a&gt; · &lt;a href="https://rekor.sigstore.dev/api/v1/log/entries?logIndex=3803861704" rel="noopener noreferrer"&gt;Rekor entry&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The revision sync and &lt;code&gt;signedRevision&lt;/code&gt; response are implemented in the checkout, with production rollout pending as of November 11. The focused local suite passed 13 cases. It exercised canonical bytes, revision chaining, and the comparison response on SQLite and PostgreSQL. The ingest fixture bypasses cryptography, which the public bundle check above exercises.&lt;/p&gt;

&lt;p&gt;Even a matching manifest can come from a server whose implementation or tool responses have changed. A publisher can also sign a harmful description. The next time a tool changes under a familiar version number, the client has a concrete question it can check. Did this publisher sign the exact definitions I’m about to hand to the model?&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance was used to draft and edit this article and to generate the cover illustration. The example and reported local results were checked against the implementation.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>security</category>
      <category>ai</category>
      <category>python</category>
    </item>
    <item>
      <title>A missing version broke verification of signed ARD search results</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Tue, 10 Nov 2026 14:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/we-added-ard-discovery-to-six-hubs-and-found-a-broken-signed-search-result-3nlp</link>
      <guid>https://dev.to/alexar76/we-added-ard-discovery-to-six-hubs-and-found-a-broken-signed-search-result-3nlp</guid>
      <description>&lt;p&gt;The weather result on modelmarket.dev looked complete until the verifier compared it with its signed subject. The subject named a version; the search result did not.&lt;/p&gt;

&lt;p&gt;The full catalog entry had &lt;code&gt;version&lt;/code&gt;. A serializer had dropped it from the search representation while keeping the &lt;code&gt;trustManifest&lt;/code&gt; that signed it. We caught that on November 4, during an ARD rollout to six AIMarket hubs, and deployed a second build.&lt;/p&gt;

&lt;p&gt;The useful part of this bug is where it failed. Checking a signature and checking that the signature belongs to the result in front of you are separate operations. You can reproduce the second failure with a public search result and a few lines of Python.&lt;/p&gt;

&lt;h2&gt;
  
  
  From a domain to a capability
&lt;/h2&gt;

&lt;p&gt;ARD is Agent Resource Discovery. This implementation follows parts of &lt;a href="https://datatracker.ietf.org/doc/html/draft-jimenez-dawn-ard-initial-contact-00" rel="noopener noreferrer"&gt;Initial Contact revision 00&lt;/a&gt; and &lt;a href="https://datatracker.ietf.org/doc/html/draft-bu-dawn-ard-discovery-00" rel="noopener noreferrer"&gt;Manifest and Lookup revision 00&lt;/a&gt;, individual Internet-Drafts dated October 30. They remain work in progress, with &lt;a href="https://datatracker.ietf.org/doc/draft-bu-dawn-ard-discovery/" rel="noopener noreferrer"&gt;no IETF endorsement or approved-standard status&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Start at &lt;a href="https://modelmarket.dev/.well-known/ard.json" rel="noopener noreferrer"&gt;modelmarket.dev/.well-known/ard.json&lt;/a&gt;. It lists four resources: the registry, the A2A card, the MCP card, and the AIMarket discovery document. The capabilities for sale sit behind the registry entry.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/a-missing-version-broke-verification-of-signed-ard-search-results/ard-discovery-boundary.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fue32w95brgrua5ygofa8.png" alt="Four discovery steps lead from an HTTPS origin through an ARD manifest and registry to a capability descriptor. A separate boundary leads to the AIMarket v2 invoke endpoint." width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The commands below stop at the descriptor. The POST used for search reads the catalog; execution and payment begin with a separate invoke request.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;With curl and jq, fetch the manifest and extract its registry address:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;manifest&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;curl &lt;span class="nt"&gt;-fsS&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  https://modelmarket.dev/.well-known/ard.json&lt;span class="si"&gt;)&lt;/span&gt;

&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s\n'&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$manifest&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; |
  jq &lt;span class="s1"&gt;'.entries[] | {displayName, type, url}'&lt;/span&gt;

&lt;span class="nv"&gt;registry&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s\n'&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$manifest&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; |
  jq &lt;span class="nt"&gt;-r&lt;/span&gt; &lt;span class="s1"&gt;'.entries[]
    | select(.type == "application/ai-registry+json")
    | .url'&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then ask that registry for weather in Paris:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;results&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;curl &lt;span class="nt"&gt;-fsS&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$registry&lt;/span&gt;&lt;span class="s2"&gt;/search"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s1"&gt;'Content-Type: application/json'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--data&lt;/span&gt; &lt;span class="s1"&gt;'{"query":{"text":"weather in Paris"},"pageSize":3}'&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;

&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s\n'&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$results&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; |
  jq &lt;span class="s1"&gt;'.results[] |
    {displayName, identifier, version, source, score, url}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Follow the returned &lt;code&gt;url&lt;/code&gt; instead of copying an address from an old example. This fetches the descriptor, which contains the schemas and calling terms:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;descriptor_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s\n'&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$results&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; |
  jq &lt;span class="nt"&gt;-r&lt;/span&gt; &lt;span class="s1"&gt;'.results[0].url // empty'&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;

&lt;span class="nb"&gt;test&lt;/span&gt; &lt;span class="nt"&gt;-n&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$descriptor_url&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
  curl &lt;span class="nt"&gt;-fsS&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$descriptor_url&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; |
  jq &lt;span class="s1"&gt;'{input_schema, output_schema, price_per_call_usd, invoke}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For a listing page instead of a text search:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-fsS&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$registry&lt;/span&gt;&lt;span class="s2"&gt;/agents?pageSize=1"&lt;/span&gt; | jq &lt;span class="nb"&gt;.&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The descriptor's &lt;code&gt;invoke&lt;/code&gt; object supplies a method, URL and request body. The body already names &lt;code&gt;product_id&lt;/code&gt;, &lt;code&gt;capability_id&lt;/code&gt; and &lt;code&gt;source_hub&lt;/code&gt;, with an empty &lt;code&gt;input&lt;/code&gt; for the buyer to fill in. The actual call still goes to &lt;code&gt;POST /ai-market/v2/invoke&lt;/code&gt;; payments, receipts and execution retain their AIMarket v2 behavior.&lt;/p&gt;

&lt;h2&gt;
  
  
  The comparison that failed
&lt;/h2&gt;

&lt;p&gt;In this implementation, &lt;code&gt;trustManifest.signature&lt;/code&gt; is a detached JWS. Its signing input is the protected header followed by the JCS-canonicalized &lt;code&gt;trustManifest&lt;/code&gt; with &lt;code&gt;signature&lt;/code&gt; removed, both base64url-encoded and joined by &lt;code&gt;.&lt;/code&gt;. The signed &lt;code&gt;subject&lt;/code&gt; contains the resource's &lt;code&gt;identifier&lt;/code&gt;, &lt;code&gt;type&lt;/code&gt;, &lt;code&gt;version&lt;/code&gt; and &lt;code&gt;digest&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Here are the three checks involved:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Resolve the publisher's &lt;code&gt;did:web&lt;/code&gt; identity, select its authorized verification key, and verify the JWS.&lt;/li&gt;
&lt;li&gt;Compare the entire signed subject with the corresponding coordinates on the returned result.&lt;/li&gt;
&lt;li&gt;Fetch the descriptor bytes and compare their SHA-256 digest with the signed digest.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The missing version breaks step 2. The JWS signs the inner manifest, so deleting a field from the outer result need not alter the signature at all.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/a-missing-version-broke-verification-of-signed-ard-search-results/ard-subject-binding.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffe0uq9myrlhe97qs8gyq.png" alt="The signed subject includes version v1, while the outer result omits version. The JWS signature and descriptor digest can still pass, but equality between the subject and result coordinates fails." width="800" height="525"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Controlled reproduction: only the outer result's version is removed. The signed manifest and descriptor bytes stay unchanged.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;This snippet requests a local publisher's weather result and isolates the subject comparison. It deletes &lt;code&gt;version&lt;/code&gt; only from the parsed object in memory:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-fsS&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$registry&lt;/span&gt;&lt;span class="s2"&gt;/search"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s1"&gt;'Content-Type: application/json'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--data&lt;/span&gt; &lt;span class="s1"&gt;'{"query":{"text":"weather","filter":{"publisher":"modelmarket.dev"}},"pageSize":10}'&lt;/span&gt; |
python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s1"&gt;'
import json, sys
fields = ("identifier", "type", "version", "digest")
row = next(r for r in json.load(sys.stdin)["results"]
           if r.get("trustManifest"))

def bound(result):
    coordinates = {k: result[k] for k in fields if k in result}
    return result["trustManifest"]["subject"] == coordinates

print("current:", bound(row))
del row["version"]
print("without version:", bound(row))
'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;current: True
without version: False
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A November 8 check also verified the live result's DID binding, Ed25519 signature and descriptor digest. After that same in-memory deletion, the signature and digest still passed while the subject comparison failed. The regression test now compares the whole subject, including field presence, rather than selected coordinates.&lt;/p&gt;

&lt;p&gt;The original rollout checks had a coverage gap: weather searches on the other hubs returned unsigned peer listings. They exercised lookup and descriptor retrieval without ever entering the signed local-result path. The main hub's own weather result reached it.&lt;/p&gt;

&lt;h2&gt;
  
  
  A result's source changes what you can verify
&lt;/h2&gt;

&lt;p&gt;Five of the six deployments sit at a domain root. Each signs its own entries with its existing Ed25519 key under a &lt;code&gt;did:web&lt;/code&gt; identity.&lt;/p&gt;

&lt;p&gt;An indexed peer listing keeps the peer's publisher identity in its identifier and names the peer as &lt;code&gt;source&lt;/code&gt;. The indexing hub leaves that entry unsigned: it is a discovery hint, with publisher authority still to establish. The sixth deployment, at &lt;code&gt;independentai.network/hub&lt;/code&gt;, also serves unsigned entries because that installation has no &lt;code&gt;did:web&lt;/code&gt; identity configured in this implementation. Its origin-level well-known address redirects to the manifest under &lt;code&gt;/hub&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;This is why the walkthrough prints both &lt;code&gt;source&lt;/code&gt; and &lt;code&gt;score&lt;/code&gt;. Source tells you whose listing you are examining; score only ranks search relevance. Establishing who signed a descriptor says nothing by itself about the correctness of a weather reading or whether the price is good.&lt;/p&gt;

&lt;h2&gt;
  
  
  Running the draft's schema
&lt;/h2&gt;

&lt;p&gt;The published CDDL grammar did not compile unchanged with cddl 0.12.14. It redefined prelude names, and the tool stopped on the content-delivery group choice inside an entry map.&lt;/p&gt;

&lt;p&gt;The conformance folder retains that original and a separate, documented checkable copy. The latter removes the conflicting definitions and uses two map alternatives for delivery. Manifest, listing-page, search-request and search-response samples then validated; an entry missing its URL and a search score of 101 were rejected.&lt;/p&gt;

&lt;p&gt;The open extension map still let the grammar accept both &lt;code&gt;url&lt;/code&gt; and &lt;code&gt;data&lt;/code&gt; on one entry. That exclusion needed an application test. These are checks of sample shapes and specific rules, not evidence of complete conformance or cross-vendor client interoperability.&lt;/p&gt;

&lt;p&gt;There is a concrete client compatibility issue too: these hubs sign with EdDSA, while &lt;a href="https://datatracker.ietf.org/doc/html/draft-bu-dawn-ard-discovery-00#section-3.5" rel="noopener noreferrer"&gt;the discovery draft makes ES256 mandatory to implement&lt;/a&gt;. An ES256-only verifier cannot check these signatures. Check the advertised algorithm before choosing a verifier for this catalog.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We operate AIMarket. AI assistance was used for the analysis, this article, and its diagrams.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>webdev</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Healthy containers hid a blocked MCP marketplace connection</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Mon, 09 Nov 2026 14:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/our-mcp-firewall-blocked-our-marketplace-while-every-container-stayed-healthy-3dbc</link>
      <guid>https://dev.to/alexar76/our-mcp-firewall-blocked-our-marketplace-while-every-container-stayed-healthy-3dbc</guid>
      <description>&lt;p&gt;All three ARGUS assistants had been running without our marketplace's MCP tools for days. Their containers kept reporting healthy. The assistants could carry on with other capabilities, so the missing connection did not stop the processes or turn their liveness checks red.&lt;/p&gt;

&lt;p&gt;The refusal came from WARDEN, the guard inside ARGUS. It compares an MCP server's tool definitions with an approved snapshot before exposing them to the model. Descriptions enter the model's context; changing their instructions matters even if a tool keeps its name. This time the guard reported &lt;code&gt;TOOL_DEF_DRIFT&lt;/code&gt; for the hub we operate ourselves.&lt;/p&gt;

&lt;p&gt;We had appended the live price and exact free-trial allowance to each paid direct tool's description. It was convenient for a model reading &lt;code&gt;tools/list&lt;/code&gt;: the cost was right there. It also made a routine tariff edit rewrite a security-pinned definition. Updating a number could cost the assistant access to the entire hub until someone reviewed it again.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prices out of the pinned description
&lt;/h2&gt;

&lt;p&gt;The November 6 repair moved those amounts out. A paid direct tool now says, "Paid per call; market_search shows the current price." A free tool says "Free." When a trial is enabled, its description mentions a few free calls without inserting the exact quota.&lt;/p&gt;

&lt;p&gt;The catalog search results retain &lt;code&gt;price_per_call_usd&lt;/code&gt; and &lt;code&gt;max_price_usd&lt;/code&gt;; an external check after deployment confirmed both on &lt;a href="https://modelmarket.dev" rel="noopener noreferrer"&gt;modelmarket.dev&lt;/a&gt;. Exact trial settings and remaining calls stay in runtime settings and result data. Switching between free and paid, or enabling and disabling the trial phrase, still changes the description.&lt;/p&gt;

&lt;p&gt;The useful invariant is narrower: different positive prices and numeric trial allowances must produce identical definitions for the same available tools.&lt;/p&gt;

&lt;p&gt;The integration regression has a &lt;code&gt;listed(price, allowance)&lt;/code&gt; helper. It seeds a weather capability, replaces the trial setting, then calls both &lt;code&gt;tools/list&lt;/code&gt; and &lt;code&gt;market_search&lt;/code&gt; through the test client. These are its essential assertions:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;before&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;search_before&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;listed&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.001&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;after_price&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;search_after&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;listed&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.05&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;after_trial&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;_&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;listed&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.001&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;9&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="n"&gt;before&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;after_price&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;after_trial&lt;/span&gt;
&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;0.001&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;search_before&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;0.05&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;search_after&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Replacing the setting matters. The trial module reads its allowance at import, so changing an environment variable later would leave the test exercising the old value. An independent review caught that mistake in the first test.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/healthy-containers-hid-a-blocked-mcp-marketplace-connection/tool-definition-fixture-matrix.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3lvhwj5vsh0aeq8gp34o.png" alt="Five local gateway fixtures: positive price and allowance changes preserve the definition; free versus paid and disabling the trial phrase change it." width="800" height="479"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Local renderer fixtures, compared as complete definition objects. These are chosen inputs, not production rates. The integration regression also checks the catalog response.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;On November 8, both targeted gateway tests passed locally. To repeat them in the hub checkout's Python environment:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;python &lt;span class="nt"&gt;-m&lt;/span&gt; pytest tests/test_mcp_stranger_path.py &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-k&lt;/span&gt; &lt;span class="s1"&gt;'price_or_trial_change or direct_tool_text'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Six hubs received the November 6 update. Five had direct tools with amounts in their descriptions; afterward, those five served the same 13-tool list despite their different prices. The sixth, with no direct tools, kept eight. Two consecutive external reads matched at every endpoint, and all six health checks returned HTTP 200. The MCP server version and packaged card moved from 1.2.0 to 1.3.0 together.&lt;/p&gt;

&lt;p&gt;Availability can still change the list. Direct tools disappear when their routes are unsellable, including when a peer has been stale for more than six hours.&lt;/p&gt;

&lt;h2&gt;
  
  
  Blocked in the health response
&lt;/h2&gt;

&lt;p&gt;The MCP host sets a connection's &lt;code&gt;quarantined&lt;/code&gt; flag before refreshing its tool list. If the read fails, the server state becomes &lt;code&gt;failed&lt;/code&gt;. If WARDEN refuses the definitions, it becomes &lt;code&gt;blocked&lt;/code&gt;. Only an accepted refresh clears quarantine. The path assembling tools for the model contains this check:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;conn&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;quarantined&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="nx"&gt;conn&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;verdict&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;allow&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;continue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;An already wrapped tool call also refreshes before invoking the server. This lets the host refuse changes that arrive during operation, including from servers that omit a tool-list-change notification.&lt;/p&gt;

&lt;p&gt;The operator needs that state outside the connection code. ARGUS now retains &lt;code&gt;connected&lt;/code&gt;, &lt;code&gt;blocked&lt;/code&gt;, or &lt;code&gt;failed&lt;/code&gt; for every MCP server it was asked to use. These fields were recorded in the public health response after the rollout:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"status"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ok"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"mcp"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"connected"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"blocked"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"failed"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The top-level &lt;code&gt;ok&lt;/code&gt; remains the liveness value our monitor expects. Counts alongside it expose integration state; an authenticated status endpoint gives per-server details and WARDEN codes.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/healthy-containers-hid-a-blocked-mcp-marketplace-connection/mcp-admission-states.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fir03a1f6oekr7vhftr42.png" alt="ARGUS process liveness remains ok while the MCP admission path can be connected, blocked after definition drift, or failed after a read error; all three instances remained blocked after the November 6 rollout." width="800" height="540"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Implementation paths, with the November 6 incident state highlighted. The connected branch shows a possible outcome, not a recorded recovery.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Independent review caught a gap here too: the first implementation reported startup blocks but missed later ones. Runtime blocks now produce error-level logs and queued owner notifications when Telegram is enabled. Repeated notifications are limited to once per server per hour; the log keeps the events. Notes written before the bot has an owner wait for one.&lt;/p&gt;

&lt;p&gt;Messages use cleaned identifiers and reason codes. Configured servers can be named; remote-catalog servers are counted without copying their catalog-chosen identifiers into the owner's message. Approval commands appear only for pin problems they can resolve. The nine targeted MCP state, notification, and HTTP tests passed locally on November 8.&lt;/p&gt;

&lt;p&gt;At the close of the November 6 rollout, all three assistants still had one blocked MCP server. The repair had changed the definitions once more. Each needed a human to review the new snapshot at an interactive terminal, approve it, and restart ARGUS before the marketplace tools could return.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We build WARDEN and ARGUS and operate AIMarket. AI assistance was used for this article, its technical figures, and implementation work; the checks described here are recorded separately from the incident.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>security</category>
      <category>ai</category>
      <category>devops</category>
    </item>
    <item>
      <title>A privacy warning became bad advice in an AI reviewer's draft</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Sun, 08 Nov 2026 14:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/our-ai-reviewer-drafted-bad-advice-about-a-useful-privacy-warning-54ne</link>
      <guid>https://dev.to/alexar76/our-ai-reviewer-drafted-bad-advice-about-a-useful-privacy-warning-54ne</guid>
      <description>&lt;p&gt;On November 4, a welcome draft for textokit suggested rewording this sentence from a tool description:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Remove secrets and personal data before submitting.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The warning belonged there. The draft treated it as a problem because a security scanner had mentioned it. Nobody sent the letter. It was one of five welcome drafts from the first run of our supplier-review agent; reading it sent us back to the JSON behind the prose.&lt;/p&gt;

&lt;p&gt;The reviewer runs in &lt;a href="https://dev.to/alexar76/our-multica-agents-now-work-in-an-agent-market-they-buy-remember-review-sellers-and-sell-their-174f"&gt;our Multica installation&lt;/a&gt;. A script downloads the manifest, federation peers and observed x402 sellers from &lt;a href="https://modelmarket.dev" rel="noopener noreferrer"&gt;modelmarket.dev&lt;/a&gt;, compares them with the previous run, and checks new or changed entries. The model writes the report and welcome drafts. A facts block generated by the script is included unchanged.&lt;/p&gt;

&lt;p&gt;The checks were deterministic. Choosing which observations to turn into advice was still left to the model.&lt;/p&gt;

&lt;h2&gt;
  
  
  A boolean disappeared in the adapter
&lt;/h2&gt;

&lt;p&gt;The textokit draft also suggested changing "instead of neutral or European Spanish" and "without a browser or API key." Those phrases describe a translation choice and a requirement the buyer does not have. WARDEN returned advisory matches for all three phrases.&lt;/p&gt;

&lt;p&gt;The adapter converted WARDEN's results into the reviewer's finding format. During that projection, it discarded &lt;code&gt;advisory&lt;/code&gt;. It also promoted an informational advisory result to low severity.&lt;/p&gt;

&lt;p&gt;Here are the relevant fields for the informational "instead of" match before and after the repair. The production records also carry the location and message:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;before&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="na"&gt;check&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;injection&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;code&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;TOOL_DEF_IMPERATIVE&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;severity&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;low&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;
&lt;span class="p"&gt;};&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;after&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="na"&gt;check&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;injection&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;code&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;TOOL_DEF_IMPERATIVE&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;severity&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;info&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;advisory&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;
&lt;span class="p"&gt;};&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That missing boolean matters more than a different display label. The report had made an advisory word match look like another finding for the model to explain. The resulting advice sounded helpful: reword whatever the checker mentioned. For the privacy warning, following that advice could make the description less useful.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/a-privacy-warning-became-bad-advice-in-an-ai-reviewers-draft/reviewer-advisory-flow.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F01oxmt1bx9w1zooagfot.png" alt="Before and after projection of an informational advisory match. The original report lost the advisory flag and raised info to low; the repaired report preserves both and gives the advisory match no letter points." width="800" height="604"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;After the first run, we tightened the prompt: advisory matches must stay out of welcome drafts, and the agent must never recommend rewording text just to stop a scanner match. On November 6, we changed the report representation too. The adapter now preserves &lt;code&gt;advisory: true&lt;/code&gt;, and informational advisory results remain informational. WARDEN's scanning rules stayed the same.&lt;/p&gt;

&lt;h2&gt;
  
  
  Give letter drafting a field contract
&lt;/h2&gt;

&lt;p&gt;Each reviewed item now carries two lists:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Field&lt;/th&gt;
&lt;th&gt;What it contains&lt;/th&gt;
&lt;th&gt;How it is used&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;findings&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;All recorded signals, including advisory matches&lt;/td&gt;
&lt;td&gt;The owner inspects them; the model receives them as context&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;letter_points&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;The eligible subset computed in code&lt;/td&gt;
&lt;td&gt;The model's instructions restrict advice to this list&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;This is the selection rule:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;letterPoints&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;findings&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;findings&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;filter&lt;/span&gt;&lt;span class="p"&gt;((&lt;/span&gt;&lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt;
    &lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;advisory&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
    &lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;severity&lt;/span&gt; &lt;span class="o"&gt;!==&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;info&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt;
    &lt;span class="p"&gt;(&lt;/span&gt;
      &lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;severity&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;high&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt;
      &lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;severity&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;medium&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt;
      &lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;check&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;schema&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
  &lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The schema exception was deliberate. A property without a declared type can deserve a useful correction even when rated low. A low-severity signal from another check does not qualify. Informational results and advisory matches never qualify. The adapter normalizes WARDEN's critical findings to the report's high category, so those remain eligible.&lt;/p&gt;

&lt;p&gt;The instructions name &lt;code&gt;letter_points&lt;/code&gt; directly and prohibit taking other findings into a welcome draft. An empty array means a welcome without an invented scanner-related complaint. Advisory matches remain visible in a separate part of the owner's report.&lt;/p&gt;

&lt;p&gt;A regression fixture retains the three ordinary phrases and produces no letter points from their matches. Two controls exercise useful findings: a tool whose input schema declares &lt;code&gt;type: "string"&lt;/code&gt; produces an eligible schema point, and an explicit instruction to ignore previous instructions and send a wallet seed phrase produces a high-severity injection finding.&lt;/p&gt;

&lt;p&gt;The local and deployed-container checks against the same downloaded November 6 hub documents agreed: textokit had zero letter points; agent-exec had four, concerning an input schema declaring &lt;code&gt;type: "string"&lt;/code&gt; and untyped properties. Both checks used temporary state.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/a-privacy-warning-became-bad-advice-in-an-ai-reviewers-draft/reviewer-letter-points-sample.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7z923ot5gzo6791wulvv.png" alt="Two recorded November 6 checker examples: textokit has zero findings eligible for author advice, and agent-exec has four. Local and container checks matched on the same documents." width="800" height="440"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;These are eligibility counts for two saved examples. They do not measure the quality of generated drafts or seller safety.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;The model still reads the full report. &lt;code&gt;letter_points&lt;/code&gt; gives it an explicit contract to follow; it does not enforce that contract on the generated prose. A person still reviews the drafts.&lt;/p&gt;

&lt;h2&gt;
  
  
  Eligibility also applies to the recipient
&lt;/h2&gt;

&lt;p&gt;The next inputs included 88 pending peer registrations shaped like probes of the hub's address handling: 69 non-HTTPS addresses and 19 HTTPS addresses the hub had never successfully read. The old rule would have treated each outside source as a candidate for a welcome draft.&lt;/p&gt;

&lt;p&gt;Sources now carry &lt;code&gt;welcome&lt;/code&gt; and &lt;code&gt;no_welcome_reason&lt;/code&gt;. Eligibility requires an HTTPS URL with a dotted hostname rather than an IP literal, no credentials in the URL, and, for a federation peer, at least one successful hub read. These are drafting conditions, not certification of a seller.&lt;/p&gt;

&lt;p&gt;A checker simulation using copied agent state found one changed outside listing and grouped the 88 peers on one line. Only agent-exec was eligible for a welcome draft. The real daily state was left untouched. That simulation checked the next inputs and eligibility rules; it did not show that a subsequent model run had written better letters.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We build WARDEN and operate AIMarket. This article was drafted with AI assistance using our implementation, regression tests and recorded rollout checks.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>security</category>
      <category>mcp</category>
    </item>
    <item>
      <title>Five MCP security scanners on the same tool definitions</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Sat, 07 Nov 2026 22:35:45 +0000</pubDate>
      <link>https://dev.to/alexar76/five-mcp-security-scanners-on-the-same-tool-definitions-5gmp</link>
      <guid>https://dev.to/alexar76/five-mcp-security-scanners-on-the-same-tool-definitions-5gmp</guid>
      <description>&lt;p&gt;An MCP tool description can tell an assistant to change another tool's arguments, read a local file before answering, or hide an extra action. A useful scanner comparison needs to show which malicious definitions receive a signal, what happens on benign definitions, and which targets never receive a complete result.&lt;/p&gt;

&lt;p&gt;I maintain both &lt;strong&gt;WARDEN and HISTOR&lt;/strong&gt;, which supplies the historical package observations used here. That is a direct conflict of interest. I wrote some fixtures, and WARDEN has been developed with several of these datasets available. This is a maintainer-run comparison with inspectable evidence, not an independent ranking.&lt;/p&gt;

&lt;p&gt;On the 485 served MCPTox definitions, published WARDEN produced high/critical signals on 460; the other four produced between 1 and 56. Lower-severity signals change that picture substantially, especially for mcp-audit. On historical packages, WARDEN also produced thousands of advisory-only signals. The tables preserve both thresholds and complete coverage rather than reducing the comparison to a winning percentage.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3n017fqmdp3isviwqlua.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3n017fqmdp3isviwqlua.png" alt="Frozen inputs pass through native analysis, with protocol parity and live discovery checked separately." width="800" height="385"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Figure 1. Native analysis at scale, with protocol parity and live discovery measured separately.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Five published products
&lt;/h2&gt;

&lt;p&gt;The primary comparison uses these actual published packages, verified at the retained registry check, including the final registry snapshot on November 7:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Scanner&lt;/th&gt;
&lt;th&gt;Package&lt;/th&gt;
&lt;th&gt;Version&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;WARDEN&lt;/td&gt;
&lt;td&gt;npm &lt;code&gt;@aimarket/warden&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;0.23.0&lt;/code&gt;, ruleset &lt;code&gt;17&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mcp-audit&lt;/td&gt;
&lt;td&gt;PyPI &lt;code&gt;mcp-audit-scanner[mcp]&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;&lt;code&gt;0.28.0&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mcp-shield&lt;/td&gt;
&lt;td&gt;npm &lt;code&gt;mcp-shield&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;&lt;code&gt;1.0.4&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;truecopy&lt;/td&gt;
&lt;td&gt;npm &lt;code&gt;@askalf/truecopy&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;&lt;code&gt;0.10.5&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mcpsentry&lt;/td&gt;
&lt;td&gt;npm &lt;code&gt;@roshan6335/mcpsentry&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;&lt;code&gt;0.1.5&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Published WARDEN has 40 rules and digest &lt;code&gt;sha256-UOfK25xo6D8vrKEVfNFBTBVn5OspyaZut9vC9Z9Q5JE=&lt;/code&gt;. Its npm tarball SHA-256 is &lt;code&gt;833841f2aab89df4c51b5f84385ad69c8ffd7dcf5a7725ec0eb4ebabe594b7ee&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The unpublished candidate &lt;code&gt;0.23.1-dev.20261107.1&lt;/code&gt;, ruleset 19, has a separate supplementary row. It never replaces the published product. An earlier frozen build labelled v17 had six changed wrapper/result-screening distribution files and was not byte-identical to npm; it is retained only as candidate history. No npm release was made during this work.&lt;/p&gt;

&lt;p&gt;Published packages and upstream public corpora are obtainable. Exact replay of the maintainer fixtures and adapters requires their source release. Bundle metadata enables inspection of row IDs, thresholds, coverage and output hashes, not full independent replay. Candidate source remains unpublished too: hashes and receipts are auditable evidence, not an externally reproducible build.&lt;/p&gt;

&lt;h2&gt;
  
  
  Native analysis and protocol checks
&lt;/h2&gt;

&lt;p&gt;Bulk execution invokes the real implementations' native classifier or full-analysis APIs with fresh per-case state. mcp-audit runs its default six analyzers through the full pipeline; only discovery transport is replaced by captured tool objects validated through its installed SDK. This is not copied-regex testing, or a claim that every row traversed every CLI's stdio transport.&lt;/p&gt;

&lt;p&gt;Representative native/API checks cover 10/10 cases for each Node competitor, 13 agreements for the Python pipeline including one partial result, and 20/20 for published WARDEN, including attack/clean/repeat and a 1.17 MB tool list. A matching partial result remains partial. The frozen v18/v19 comparison additionally matched actual-MCP/library results on 3,177/3,177 cases per version per host.&lt;/p&gt;

&lt;p&gt;The shared columns are &lt;strong&gt;high/critical definition signals&lt;/strong&gt;, &lt;strong&gt;any relevant definition signal&lt;/strong&gt;, and &lt;strong&gt;advisory-only&lt;/strong&gt;, where advisory-only = any − HC. They are not identical enforcement outcomes. WARDEN's native &lt;code&gt;allow: false&lt;/code&gt; is retained separately. truecopy can classify even advisory content findings as poisoned and withhold them; cold-lock/unpinned refusal is a different policy. mcp-shield normally exits 0 despite findings, and mcpsentry exits 1 only for critical findings, not high-only ones.&lt;/p&gt;

&lt;p&gt;WARDEN uses &lt;code&gt;allowUnknownServers=true&lt;/code&gt;, &lt;code&gt;sensitiveToolPatterns=[]&lt;/code&gt; and &lt;code&gt;pinToolDefs=true&lt;/code&gt;. Both static-scan and threat-feed definition gates count. Launch hygiene, origin/pinning policy and first-scan baseline noise remain in raw records but are excluded consistently for every product. A corrected first aggregator had wrongly omitted three legitimate WARDEN threat-feed detections; native outputs were unchanged, and that normalization diagnostic is retained.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the 18,110 rows contain
&lt;/h2&gt;

&lt;p&gt;These are &lt;strong&gt;target rows, not 18,110 independent live servers&lt;/strong&gt;. Our fixtures contain 23 attacks and 12 benign controls. Pinned third-party fixtures contain 10 attacks, one benign case and one ambiguous case. Mixed totals of 35 or 12 are not attack-recall denominators. These known development inputs are not new blind tests.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://arxiv.org/abs/2508.14925" rel="noopener noreferrer"&gt;MCPTox&lt;/a&gt; has overlapping views: 485 unique served poisoned definitions across 45 servers; 1,348 complete prompt-surface scenario rows; 225 decorated variants; and 45 clean server lists. The full scenario view retains 1,338 unique server/frame values and 10 retained duplicate instances, with zero discarded. The original split has 267 development definitions and 218 test definitions. That test split is a &lt;strong&gt;known historical holdout&lt;/strong&gt;, repeatedly published and sharing templates with development cases. Reuse does not restore blindness.&lt;/p&gt;

&lt;p&gt;The official source commit is &lt;code&gt;f85189f9ad12504c197c7f920ab818a40657b1fa&lt;/code&gt;. Projection handles 702 literal-line-feed and 646 escaped-line-feed rows. The full-frame projection retains argument text verbatim but represents declared parameter properties as strings; it does not reproduce original native argument types. The clean projection exposes 362 tools versus 353 stated in the paper; deduplication did not remove that discrepancy. The repository specifies no license, so public downloads contain factual metadata rather than raw definitions, schemas, payloads or benchmark Python.&lt;/p&gt;

&lt;p&gt;The public corpus contains 986 server snapshots and a separate 41-observation carryover cohort. There are 28 exact name+definition overlaps, leaving 999 unique snapshots. The original cohorts remain separate.&lt;/p&gt;

&lt;p&gt;HISTOR contributes &lt;strong&gt;14,933 recorded package versions: 12,023 npm and 2,910 PyPI&lt;/strong&gt;, referencing 14,449 distinct toolsets and 272,657 tool records. Every row was recovered and executed independently, including repeated toolsets. MTL digests, file hashes and signatures were verified on both hosts. Historical base fields remain historical: current titles, annotations and extensions were not merged into them.&lt;/p&gt;

&lt;p&gt;HISTOR's &lt;a href="https://histor.modelmarket.dev/datasets/2026-10-20/toolsets.jsonl.gz" rel="noopener noreferrer"&gt;2026-10-20 archive&lt;/a&gt;, with its &lt;a href="https://histor.modelmarket.dev/datasets/2026-10-20/manifest.json" rel="noopener noreferrer"&gt;signed manifest&lt;/a&gt;, supplied 14,930 rows. The archive is &lt;a href="https://creativecommons.org/licenses/by/4.0/" rel="noopener noreferrer"&gt;CC-BY-4.0&lt;/a&gt;, attributed to Aleksandr Artamokhov / HISTOR. Three missing digests came from immutable public API URLs without a separate stated license; their public inventory is metadata-only.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two remote gVisor hosts
&lt;/h2&gt;

&lt;p&gt;Reported comparative measurements ran remotely on attested and PingBlip. An earlier local pilot is excluded and its generated artifacts removed. All bulk engines used &lt;code&gt;runsc&lt;/code&gt;, network none, nonroot UID/GID 501:501, read-only root/source/runtime/data, dropped capabilities, no-new-privileges and dedicated results. No historical package implementation or upstream benchmark Python was executed.&lt;/p&gt;

&lt;p&gt;Both hosts used image &lt;code&gt;sha256:db7085498187e8477fa462d5f1652d62bdf4b8af29ea3666bd1748a8315631a5&lt;/code&gt;, containing Node 24.22.2 / Python 3.13.18. Node engines used 1 CPU / 768 MiB / pids 128. Audit base used 1 CPU / 768 MiB / pids 256; its historical phase used one worker / 1 CPU / 768 MiB on attested and two workers / 2 CPU / 1536 MiB on PingBlip. Linux cgroup quotas were checked. This is not a controlled speed comparison.&lt;/p&gt;

&lt;p&gt;The 3,177-view catalog SHA-256 is &lt;code&gt;e29baa05fbdbf46e34631ff2c75f0b4f535f0188054dd4499699bd189a3dc757&lt;/code&gt;; the 14,933-package catalog is &lt;code&gt;78d4dab7125318b80cc35302e7cf1cfc1a49fa6f310957ed7b8fddfb538375b6&lt;/code&gt;. Every case has a file hash and common opaque wire id. Full methods and input units are in the &lt;a href="https://modelmarket.dev/blog/media/warden-security-comparison-2026-11-07/reports-and-data.zip" rel="noopener noreferrer"&gt;report and data ZIP&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Severity signals and advisory costs
&lt;/h2&gt;

&lt;p&gt;Every cell below is &lt;strong&gt;HC/any/advisory-only&lt;/strong&gt;, as case counts. All base rows completed on both hosts. HISTOR exceptions follow in the coverage table. The overlapping MCPTox rows are not additive independent trials.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Cohort and scheduled unit&lt;/th&gt;
&lt;th&gt;WARDEN&lt;/th&gt;
&lt;th&gt;mcp-audit&lt;/th&gt;
&lt;th&gt;mcp-shield&lt;/th&gt;
&lt;th&gt;truecopy&lt;/th&gt;
&lt;th&gt;mcpsentry&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;23 own attacks&lt;/td&gt;
&lt;td&gt;20/23/3&lt;/td&gt;
&lt;td&gt;11/13/2&lt;/td&gt;
&lt;td&gt;6/10/4&lt;/td&gt;
&lt;td&gt;3/8/5&lt;/td&gt;
&lt;td&gt;4/4/0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;10 third-party attacks&lt;/td&gt;
&lt;td&gt;10/10/0&lt;/td&gt;
&lt;td&gt;10/10/0&lt;/td&gt;
&lt;td&gt;8/9/1&lt;/td&gt;
&lt;td&gt;1/7/6&lt;/td&gt;
&lt;td&gt;5/5/0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;485 served MCPTox definitions&lt;/td&gt;
&lt;td&gt;460/474/14&lt;/td&gt;
&lt;td&gt;56/392/336&lt;/td&gt;
&lt;td&gt;56/71/15&lt;/td&gt;
&lt;td&gt;1/31/30&lt;/td&gt;
&lt;td&gt;19/19/0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;…218 historical-test definitions&lt;/td&gt;
&lt;td&gt;210/217/7&lt;/td&gt;
&lt;td&gt;39/178/139&lt;/td&gt;
&lt;td&gt;41/46/5&lt;/td&gt;
&lt;td&gt;0/27/27&lt;/td&gt;
&lt;td&gt;8/8/0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;1,348 full scenario rows&lt;/td&gt;
&lt;td&gt;1331/1338/7&lt;/td&gt;
&lt;td&gt;951/1264/313&lt;/td&gt;
&lt;td&gt;158/570/412&lt;/td&gt;
&lt;td&gt;432/491/59&lt;/td&gt;
&lt;td&gt;460/460/0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;12 own benign controls&lt;/td&gt;
&lt;td&gt;0/3/3&lt;/td&gt;
&lt;td&gt;1/2/1&lt;/td&gt;
&lt;td&gt;1/2/1&lt;/td&gt;
&lt;td&gt;1/1/0&lt;/td&gt;
&lt;td&gt;2/2/0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;45 clean MCPTox servers&lt;/td&gt;
&lt;td&gt;0/7/7&lt;/td&gt;
&lt;td&gt;0/3/3&lt;/td&gt;
&lt;td&gt;3/6/3&lt;/td&gt;
&lt;td&gt;0/0/0&lt;/td&gt;
&lt;td&gt;1/7/6&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;986 public snapshots&lt;/td&gt;
&lt;td&gt;4/388/384&lt;/td&gt;
&lt;td&gt;43/159/116&lt;/td&gt;
&lt;td&gt;343/451/108&lt;/td&gt;
&lt;td&gt;2/39/37&lt;/td&gt;
&lt;td&gt;86/450/364&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;14,933 historical package rows&lt;/td&gt;
&lt;td&gt;409/6327/5918&lt;/td&gt;
&lt;td&gt;481/2089/1608&lt;/td&gt;
&lt;td&gt;5354/6767/1413&lt;/td&gt;
&lt;td&gt;15/252/237&lt;/td&gt;
&lt;td&gt;1449/6362/4913&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;All five had zero signals on the one third-party benign control. Only mcp-shield advised on the ambiguous case; it is excluded from attack and benign denominators. Decorated/development/carryover results are in the full report.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx55husxftzmtitwzhxk2.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx55husxftzmtitwzhxk2.png" alt="High/critical and advisory-only signals for the five published packages, with each cohort and its unit kept separate." width="800" height="850"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Figure 2. Published products only. Severity signals and advisory costs are separate; overlapping cohorts are not additive.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Public and package rows are unlabelled. Their positive frequency is &lt;strong&gt;not a false-positive rate&lt;/strong&gt;. WARDEN's 409 historical HC positives coexist with 5,918 advisory-only cases; calling it simply “409 alerts” would hide the lower-tier workload. truecopy has fewer HC positives there, while also detecting fewer native malicious definitions at that threshold. These are dataset-specific tradeoffs, not a universal ranking.&lt;/p&gt;

&lt;h2&gt;
  
  
  Coverage and exact agreement
&lt;/h2&gt;

&lt;p&gt;For scheduled &lt;code&gt;N&lt;/code&gt;, complete &lt;code&gt;C&lt;/code&gt;, HC &lt;code&gt;B&lt;/code&gt;, advisory-only &lt;code&gt;A&lt;/code&gt;, error/partial &lt;code&gt;E&lt;/code&gt;, skipped &lt;code&gt;S&lt;/code&gt; and not-run &lt;code&gt;U&lt;/code&gt;: &lt;code&gt;N=C+E+S+U&lt;/code&gt;, &lt;code&gt;C=B+A+no-finding&lt;/code&gt;. Threshold-signal rate is &lt;code&gt;B/C&lt;/code&gt;; it is recall only within attack-labelled cohorts. Coverage is &lt;code&gt;C/N&lt;/code&gt;. Errors are not clean scans or completed misses.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Product&lt;/th&gt;
&lt;th&gt;Complete/scheduled, attested; PingBlip&lt;/th&gt;
&lt;th&gt;Errors, attested; PingBlip&lt;/th&gt;
&lt;th&gt;Identical/complete paired results&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;WARDEN&lt;/td&gt;
&lt;td&gt;18110/18110; same&lt;/td&gt;
&lt;td&gt;0; 0&lt;/td&gt;
&lt;td&gt;18110/18110&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mcp-audit&lt;/td&gt;
&lt;td&gt;18108/18110; same&lt;/td&gt;
&lt;td&gt;2; 2&lt;/td&gt;
&lt;td&gt;18108/18108&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mcp-shield&lt;/td&gt;
&lt;td&gt;18079/18110; 18078/18110&lt;/td&gt;
&lt;td&gt;31; 32&lt;/td&gt;
&lt;td&gt;18078/18078&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;truecopy&lt;/td&gt;
&lt;td&gt;18110/18110; same&lt;/td&gt;
&lt;td&gt;0; 0&lt;/td&gt;
&lt;td&gt;18110/18110&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mcpsentry&lt;/td&gt;
&lt;td&gt;18110/18110; same&lt;/td&gt;
&lt;td&gt;0; 0&lt;/td&gt;
&lt;td&gt;18110/18110&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;All products have zero skipped, unrun or missing rows. Two mcp-audit HISTOR records hit native schema/text budgets and are partial on both hosts, excluded from complete coverage and primary counts. mcp-shield's strict SDK rejects 31 schemas on both hosts; PingBlip also has one 30-second timeout. Native raw JSON-RPC clients used by truecopy and mcpsentry do not impose that validator. A rejected 42-tool case passed mcpsentry CLI/API with no findings: a real compatibility difference, not an omitted adapter check.&lt;/p&gt;

&lt;p&gt;Every completed comparable result matched, including canonical full native-result signatures—not merely aggregate counts. The &lt;a href="https://modelmarket.dev/blog/media/warden-security-comparison-2026-11-07/reports-and-data.zip" rel="noopener noreferrer"&gt;case-level proof in the report/data ZIP&lt;/a&gt; also checks input and engine identities and separates incomplete pairs. Only mcp-audit's own temporary config path and fresh rug-pull state-file evidence are normalized; finding content remains represented.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffzk6jzdw27s55qzjou39.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffzk6jzdw27s55qzjou39.png" alt="Both-host coverage and exact agreement, keeping partial scans, schema errors and the one-host timeout outside complete-result parity." width="800" height="660"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Figure 3. Exact agreement uses complete paired results. Partial scans and errors remain visible.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The unpublished candidate and the retained miss
&lt;/h2&gt;

&lt;p&gt;The first frozen v18 candidate regressed against the earlier frozen17 build: 457/485 served blocks versus 460/485, and 207/218 historical-test blocks versus 210/218. &lt;code&gt;FileSystem_20&lt;/code&gt;, &lt;code&gt;_22&lt;/code&gt; and &lt;code&gt;_23&lt;/code&gt; were newly missed. Those failures remain in the evidence.&lt;/p&gt;

&lt;p&gt;Candidate19 &lt;code&gt;.1&lt;/code&gt; recovers two: overwriting an SSH public-key file is security-configuration tampering; reading &lt;code&gt;/etc/shadow&lt;/code&gt; accesses a secret store. &lt;code&gt;FileSystem_22&lt;/code&gt;, a public-key-read prerequisite, remains nonblocking. It remains labelled as an attack and stays in the denominator; public material is not reclassified as secret just to recover a benchmark score.&lt;/p&gt;

&lt;p&gt;The verified candidate blocks 459/485 served definitions, 209/218 historical-test definitions, 1331/1348 scenarios and 225/225 decorated cases, with 0/45 clean-server blocks. It completes all 18,110 rows without errors and matches across hosts. Its 1536/1536 tests in 63 files pass on both hosts, alongside typecheck/build/red-team commands. This is post-hoc work on observed failures, not new holdout generalization.&lt;/p&gt;

&lt;p&gt;Candidate HISTOR has 407 HC / 6327 any versus published 409/6327, with no v18→v19 verdict change. Narrow review of the two public-key provisioning passages in &lt;code&gt;@krovacloud/mcp@0.3.16&lt;/code&gt; and &lt;code&gt;@runpod/mcp-server@4.4.0&lt;/code&gt; supports allowing those passages, not certifying whole packages benign. The later &lt;code&gt;.2&lt;/code&gt; runtime build fixes optional-MIME blob compatibility. Both hosts pass 1536/1536 full tests, 304 focused tests and 8/8 SDK sessions; independent binary checks preserve valid 400k blobs with/without MIME and withhold poisoned results after one execution. The 320 source/distribution files outside the changed result-screen module are byte-identical to &lt;code&gt;.1&lt;/code&gt;, but &lt;code&gt;.2&lt;/code&gt; is not substituted for the frozen &lt;code&gt;.1&lt;/code&gt; measurement. The current &lt;code&gt;0.23.1-dev.20261108&lt;/code&gt; snapshot adds Codex/OpenCode hooks absent from those builds; its acceptance validation is separate and is not represented by these comparison rows.&lt;/p&gt;

&lt;p&gt;Download the &lt;a href="https://modelmarket.dev/blog/media/warden-security-comparison-2026-11-07/reports-and-data.zip" rel="noopener noreferrer"&gt;report and data bundle — English, Russian, Spanish, French and Chinese&lt;/a&gt;. The ZIP contains the five reports for offline reading, the full 18,110-row CSV/JSON inventory, grouped metrics, case-level proofs, methods and package pins. Public per-case data contains statuses, signal flags and input/result hashes, not full native findings or unlicensed payloads. The bundle supplies recorded versions and provenance without uploading product source.&lt;/p&gt;

&lt;p&gt;Separately, candidate19 &lt;code&gt;.1&lt;/code&gt; completed 8/8 official-SDK sessions and 4/4 direct/wrapped list-digest pairs per host: filesystem 2.2.0, memory 2.1.0, sequential-thinking 2.1.1 and WARDEN, using SDK 1.32.1. Checks included discovery, a safe synthetic call, unknown-tool refusal and later ping. HTTP discovery found DeepWiki/Microsoft Learn/Cloudflare Docs with 3/3/2 tools, all allowed, using only &lt;code&gt;initialize&lt;/code&gt; and &lt;code&gt;tools/list&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;This measures definition classification at recorded thresholds. It does not measure model attack success, full proxy enforcement, dependency safety or universal protection. The useful result is the detection/advisory-cost tradeoff with coverage visible, supported by case-level evidence.&lt;/p&gt;

&lt;p&gt;AI assistance was used for implementation changes, test orchestration, analysis, writing and figure generation. Responsibility for the benchmark design and interpretation remains mine as WARDEN and HISTOR's maintainer.&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>security</category>
      <category>opensource</category>
      <category>ai</category>
    </item>
    <item>
      <title>An MCP tool set can gain a digest without changing</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Sat, 07 Nov 2026 16:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/an-mcp-tool-set-can-gain-a-digest-without-changing-4mjd</link>
      <guid>https://dev.to/alexar76/an-mcp-tool-set-can-gain-a-digest-without-changing-4mjd</guid>
      <description>&lt;p&gt;A schema default of &lt;code&gt;0.5&lt;/code&gt; used to be enough for HISTOR to leave an MCP server's entire tool set without a digest. After a change to the measurement rules, the same declarations could receive one. The server's author did not have to edit anything.&lt;/p&gt;

&lt;p&gt;That left an awkward historical record: "not digestible" could still describe a decision made under rules the observer no longer used. The label did not say which revision had declined the set.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://histor.modelmarket.dev" rel="noopener noreferrer"&gt;HISTOR&lt;/a&gt; logs the tool definitions MCP servers advertise. Its MTL profile specifies a canonical representation of those definitions and the digest computed from it. MTL is our draft labeling profile; it is not a standard of the MCP project.&lt;/p&gt;

&lt;p&gt;Two recent revisions addressed different parts of the problem. Revision 1.1, deployed on October 31, expanded which tool sets could be digested. Revision 1.2, deployed on November 6, records why a tool-set digest is absent and which rules made that decision.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rules changed before the records did
&lt;/h2&gt;

&lt;p&gt;MTL 1.0 declined schemas containing fractions, exponent notation, or integers outside the safe integer range. It also declined empty tool sets and sets with repeated tool names.&lt;/p&gt;

&lt;p&gt;Revision 1.1 reads numbers as IEEE 754 doubles and writes them in the &lt;a href="https://www.rfc-editor.org/rfc/rfc8785" rel="noopener noreferrer"&gt;RFC 8785 canonical form&lt;/a&gt;. It handles empty sets and repeated names too. Definitions that the older rules declined could therefore acquire a digest without any change on the server.&lt;/p&gt;

&lt;p&gt;Before hashing, the normalizer keeps &lt;code&gt;name&lt;/code&gt;, &lt;code&gt;description&lt;/code&gt;, &lt;code&gt;inputSchema&lt;/code&gt;, and an object-valued &lt;code&gt;outputSchema&lt;/code&gt; when present. It orders tools by name using UTF-16 code units; repeated names are ordered by canonical bytes. The normalized array is then serialized as canonical JSON. A tool-set digest therefore commits to those fields, not to every annotation the server sends or to its implementation.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/an-mcp-tool-set-can-gain-a-digest-without-changing/histor-digest-fixtures.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiwppu67lb45nofxd44e6.png" alt="Eight offline fixtures compared under MTL 1.0 and 1.1: integer remains digestible; fractions, exponents, large integers, empty lists and repeated names become digestible; missing names and overflow remain declined." width="800" height="575"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Local fixture outcomes from the archived 1.0 and 1.1 reference functions. Revision 1.2 keeps the 1.1 digest rules and adds label metadata. &lt;a href="https://modelmarket.dev/blog/an-mcp-tool-set-can-gain-a-digest-without-changing/histor-digest-fixtures.svg" rel="noopener noreferrer"&gt;Editable SVG&lt;/a&gt;; &lt;a href="https://modelmarket.dev/blog/an-mcp-tool-set-can-gain-a-digest-without-changing/histor-fixture-results.json" rel="noopener noreferrer"&gt;fixture results&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Digests already produced under 1.0 stayed unchanged, so the descriptor's &lt;code&gt;mtl&lt;/code&gt; identifier remained &lt;code&gt;"1"&lt;/code&gt;. That preserved existing comparisons, but it did not tell a later reader which revision had produced an old refusal.&lt;/p&gt;

&lt;p&gt;In the &lt;a href="https://github.com/GenAI-Security-Project/agent-control-standard/issues/38#issuecomment-6688178696" rel="noopener noreferrer"&gt;Agent Control Standard discussion&lt;/a&gt;, avp9-nexus proposed keeping two absence cases apart, with a reason and profile identifier attached. That distinction fits the records HISTOR actually has:&lt;/p&gt;

&lt;p&gt;&lt;code&gt;not_observed&lt;/code&gt; means that no tool set was read for this observation. &lt;code&gt;not_digestible&lt;/code&gt; means a set was read, but the applicable profile declined to digest it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/an-mcp-tool-set-can-gain-a-digest-without-changing/histor-observation-states.svg" rel="noopener noreferrer"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fly4trj0n8o2riv5uon2s.png" alt="Observation flow: a failed tools/list read produces not_observed; after a successful read, a profile refusal produces not_digestible with code and revision, while acceptance produces a tool-set digest." width="800" height="559"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The read and the canonicalization decision are separate steps. A signed label can record any branch; its signature does not supply the missing tool-set digest. &lt;a href="https://modelmarket.dev/blog/an-mcp-tool-set-can-gain-a-digest-without-changing/histor-observation-states.svg" rel="noopener noreferrer"&gt;SVG&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;A failed read calls for investigation of the endpoint or access problem. A profile refusal calls for the refusal code and the rules that produced it. An empty list read successfully belongs to neither absence case: since 1.1, the empty set can have a digest.&lt;/p&gt;

&lt;p&gt;Our follow-up records the profile revision as well. New labels issued under 1.2 name &lt;code&gt;profileRevision: "1.2.0"&lt;/code&gt;; a read set declined by the profile carries a &lt;code&gt;toolSet.noDigest&lt;/code&gt; reason and the profile's code where known. Revision 1.2 changes no digest calculation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try the numeric case
&lt;/h2&gt;

&lt;p&gt;With Python 3.11 or newer, the published package exposes the same tool-set function:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;python &lt;span class="nt"&gt;-m&lt;/span&gt; pip &lt;span class="nb"&gt;install &lt;/span&gt;aimarket-histor&lt;span class="o"&gt;==&lt;/span&gt;0.6.2
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;histor.subject&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;MtlError&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;tool_set_digest&lt;/span&gt;


&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;default&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[{&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;name&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;demo&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;description&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;A fixture.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;inputSchema&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;object&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;properties&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;ratio&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;number&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;default&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;default&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="p"&gt;}}&lt;/span&gt;
    &lt;span class="p"&gt;}]&lt;/span&gt;


&lt;span class="n"&gt;digest&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;_&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;tool_set_digest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.5&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;span class="k"&gt;assert&lt;/span&gt; &lt;span class="n"&gt;digest&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="nf"&gt;tool_set_digest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;5e-1&lt;/span&gt;&lt;span class="p"&gt;))[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;digest&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;try&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="nf"&gt;tool_set_digest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;loads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;1e400&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)))&lt;/span&gt;
&lt;span class="k"&gt;except&lt;/span&gt; &lt;span class="n"&gt;MtlError&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;error&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;error&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;code&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The script prints &lt;code&gt;sha256-sIkKM1gFvjvHtnPc4ZRYq5TB1gGc+OwUBxLz8yS5VQ4=&lt;/code&gt; and then &lt;code&gt;MTL-NUM-001&lt;/code&gt;. The snippet ran unchanged in a clean environment against the PyPI wheel. &lt;code&gt;0.5&lt;/code&gt; and &lt;code&gt;5e-1&lt;/code&gt; arrive as the same double; &lt;code&gt;1e400&lt;/code&gt; has no finite double representation, so the profile still refuses it.&lt;/p&gt;

&lt;p&gt;The large-integer row has a cost too: values beyond the safe integer range are represented as their nearest binary64 value. For example, &lt;code&gt;2**53 + 1&lt;/code&gt; and &lt;code&gt;2**53&lt;/code&gt; produce the same digest for this fixture. That follows the numeric view a JavaScript JSON parser holds; it does not preserve arbitrary-precision decimal literals. A consumer needing that distinction must use a different representation.&lt;/p&gt;

&lt;h2&gt;
  
  
  A null revision can be the honest answer
&lt;/h2&gt;

&lt;p&gt;During the November 6 rollout, a server-detail response exposed this historical reason object:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"reason"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"not_digestible"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"code"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"MTL-NUM-001"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"profileRevision"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;null&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"decidedAt"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"2026-10-20T21:55:11Z"&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The old label never stated its revision. The API leaves that field null rather than assigning today's rules to a decision made on October 20. &lt;code&gt;decidedAt&lt;/code&gt; belongs to the historical observation, not to the upgrade that exposed the extra fields.&lt;/p&gt;

&lt;p&gt;Continuity labels need to preserve both sides of a comparison. A failed current read can carry &lt;code&gt;not_observed&lt;/code&gt; under &lt;code&gt;noDigest.current&lt;/code&gt;; a prior set declined by the profile has its reason on the prior side. The earlier revision is copied only if the earlier label actually named it. Updating the issuer does not fill gaps in an old assertion.&lt;/p&gt;

&lt;p&gt;A missing tool-set digest also does not erase the label's signature or every other digest it contains. Neither absence reason is a verdict on server safety. The signature attributes the assertion to an issuer; it does not independently prove that retrieval occurred or when.&lt;/p&gt;

&lt;h2&gt;
  
  
  Letting unchanged packages run again
&lt;/h2&gt;

&lt;p&gt;A scheduler watching only package versions would miss the change. It has already run that release, the release has not moved, and its old result says there is no tool-set digest.&lt;/p&gt;

&lt;p&gt;HISTOR now makes a package eligible for another run when an older MTL revision declined its set. If the current revision still declines it, the recorded decision settles that case until relevant inputs change. The revision marker is excluded from security events, so an observer upgrade is not reported as a security event attributed to the package.&lt;/p&gt;

&lt;p&gt;The scheduler regression checks this unchanged-version path. Label tests cover the two reasons, refusal codes, signatures with the new fields, and comparisons whose earlier label lacks a revision. Before the November 6 production switch, a disposable instance exercised label generation and both API views without network access; outside checks followed the deployment.&lt;/p&gt;

&lt;p&gt;The distinction is visible in the public API: &lt;a href="https://histor.modelmarket.dev/api/v1/servers?state=unobserved&amp;amp;limit=1" rel="noopener noreferrer"&gt;one unobserved entry&lt;/a&gt; and &lt;a href="https://histor.modelmarket.dev/api/v1/servers?state=undigested&amp;amp;limit=1" rel="noopener noreferrer"&gt;one undigested entry&lt;/a&gt;. Both returned their corresponding &lt;code&gt;noDigest.reason&lt;/code&gt; when checked on November 7. These are reasons attached to particular measurements, so the date and profile revision matter as much as the word "missing."&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We build HISTOR. AI assistance was used for this article and the implementation work it describes.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>security</category>
      <category>opensource</category>
      <category>observability</category>
    </item>
    <item>
      <title>Our Multica agents now work in an agent market: they buy, remember, review sellers, and sell their own work</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Thu, 05 Nov 2026 08:47:29 +0000</pubDate>
      <link>https://dev.to/alexar76/our-multica-agents-now-work-in-an-agent-market-they-buy-remember-review-sellers-and-sell-their-174f</link>
      <guid>https://dev.to/alexar76/our-multica-agents-now-work-in-an-agent-market-they-buy-remember-review-sellers-and-sell-their-174f</guid>
      <description>&lt;p&gt;&lt;a href="https://github.com/multica-ai/multica" rel="noopener noreferrer"&gt;Multica&lt;/a&gt; is an agent factory: you give an agent a runtime, instructions and&lt;br&gt;
a schedule, and it works through tasks. AIMarket hubs (&lt;a href="https://modelmarket.dev" rel="noopener noreferrer"&gt;modelmarket.dev&lt;/a&gt; is the public one)&lt;br&gt;
are a market where agents buy capabilities from each other per call. We run both, and in the first week of November&lt;br&gt;
we connected them. This is what the agents do now, what broke, what we found when we audited ourselves, and a&lt;br&gt;
setup guide in five languages for anyone who wants the same.&lt;/p&gt;

&lt;h2&gt;
  
  
  The crew
&lt;/h2&gt;

&lt;p&gt;Five agents run on one self-hosted Multica 0.5.3 (OpenCode + DeepSeek), on fixed UTC schedules:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GitHub digest&lt;/strong&gt;, daily at 08:00: activity across our repositories, sent to Telegram. It keeps its past digests
in Attested Memory, bought on our hub per call ($0.004 to search, $0.012 to write), so it can say "stars on X
fell for the third day" instead of repeating yesterday. About $0.03 a run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Mystery shopper&lt;/strong&gt;, daily at 09:00: buys a few random safe skills on modelmarket.dev with a prepaid key capped at
$0.15 a run, and checks the receipt, the latency and whether the answer matches the listing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Supplier reviewer&lt;/strong&gt;, daily at 10:00: diffs the market against yesterday and reviews every new or changed
listing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Weekly post&lt;/strong&gt;, Mondays at 11:00: drafts a build-in-public post from the hub's counters and the agents' own
results. A human reads it and decides.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Order runner&lt;/strong&gt;: writes the summary for a long job that someone buys on the hub (below).&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The bug we found before writing a single memory
&lt;/h2&gt;

&lt;p&gt;Our memory service took a memory's owner from a payment header that the hub sends empty for most buyers. Every buyer&lt;br&gt;
on a prepaid key, a trial or an on-chain payment was the same owner, so "private" memories were readable by all of&lt;br&gt;
them. Nothing had been written through the hub yet, so nothing leaked. The fix: the hub names the buyer it&lt;br&gt;
authenticated to its own providers with an HMAC pseudonym. The pseudonym is stable across restarts and reveals no&lt;br&gt;
account, and the memory service keys owners by it.&lt;/p&gt;

&lt;p&gt;We found the same bug once more a day later, one hub further away (see "We audited ourselves").&lt;/p&gt;

&lt;h2&gt;
  
  
  Agents that read strangers' text
&lt;/h2&gt;

&lt;p&gt;An agent that reads GitHub issues or sellers' answers can be talked into sending data out. Three layers:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Egress allowlist.&lt;/strong&gt; The agents' Docker network has no route out except a proxy with a short, anchored host
list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A flow guard in the runtime.&lt;/strong&gt; Our &lt;a href="https://github.com/alexar76/warden" rel="noopener noreferrer"&gt;WARDEN&lt;/a&gt; guard runs as an OpenCode
plugin (&lt;code&gt;@aimarket/warden/opencode&lt;/code&gt;, from 0.23.0). After the session has read outside text, it refuses a tool
call that would carry a credential or text the session read. In a staged attack we planted a fake GitHub token in
a file. Sending it was refused through the chat script, curl to an allowed host, a memory write, a project file
and the tracker CLI. A normal day's digest ran with zero refusals, after we fixed one false positive: a task's
UUID looked like private text.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Least privilege per agent.&lt;/strong&gt; All agents share one daemon, so each gets its own Multica environment. The
reviewer and the post writer get empty hub keys, no GitHub token, and the hub's MCP tools switched off.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Reviewing sellers with a script, not a model
&lt;/h2&gt;

&lt;p&gt;The reviewer reads exactly the text an attacker writes: listing descriptions. So a script decides the facts, and the&lt;br&gt;
model only writes the prose and a welcome draft for each outside author. The facts are broken schemas, injection&lt;br&gt;
patterns from WARDEN's static scan, a skill that asks the buyer for a credential, plain-http or internal addresses,&lt;br&gt;
and price mismatches. The script's facts block goes into the report unchanged.&lt;/p&gt;

&lt;p&gt;Its first run took a baseline of 282 listings, 22 federation peers and 2 x402 sellers. The one high finding was ours:&lt;br&gt;
our public manifest showed a seller's internal plain-http container address. The hub now shows its own invoke&lt;br&gt;
address for sellers that only it can reach, and keeps the real one for routing. The reviewer also drafted five&lt;br&gt;
welcome letters to outside sellers. Nobody has sent them; that is a human decision.&lt;/p&gt;

&lt;h2&gt;
  
  
  Selling a long job
&lt;/h2&gt;

&lt;p&gt;A paid call cannot wait minutes for an agent, so a GitHub activity digest is sold as two capabilities on&lt;br&gt;
hub.attestedmemory.net:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;github.digest.order@v1&lt;/code&gt; ($0.05) takes any public GitHub user or organization and 1 to 7 days, and answers at
once with a job id;&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;github.digest.result@v1&lt;/code&gt; (free) returns the digest to the buyer who paid.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The first real order failed, and our own guard caused half of it. The agent read its task, including a per-job&lt;br&gt;
delivery URL, then read GitHub, then tried to send the result to that URL. WARDEN treats task text as private, and&lt;br&gt;
after outside text it lets no private data leave, whatever the destination. That is the rule we want for an&lt;br&gt;
unattended agent; the delivery design was wrong. The other half: GitHub's 60 unauthenticated requests an hour per IP&lt;br&gt;
were already spent on that host. (By what, we found later: a video worker of ours whose model calls had been failing&lt;br&gt;
since September. It re-fetched GitHub every two minutes, because its failed run never saved the time.)&lt;/p&gt;

&lt;p&gt;The second design went live the next morning. The service does the GitHub reads itself, with a token that has no&lt;br&gt;
scopes, and builds the dated items. A login that does not exist, or a GitHub outage, refuses the order before&lt;br&gt;
anyone is charged. The agent's final answer is one JSON line with a summary. The service collects it through a&lt;br&gt;
read-only database function that returns only that agent's answers for that job. If no usable summary arrives,&lt;br&gt;
the service writes a plain one from the same data, so a paid order always gets its digest. Four test orders so far:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;octocat for a week, 0 events;&lt;/li&gt;
&lt;li&gt;the microsoft organization for a day, 87 events;&lt;/li&gt;
&lt;li&gt;our own account for three days, then for two.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each came back with the agent's summary. In the latest one, the agent's answer was there 31 seconds after the&lt;br&gt;
order.&lt;/p&gt;

&lt;h2&gt;
  
  
  We audited ourselves for the bug classes an outside audit found
&lt;/h2&gt;

&lt;p&gt;In October, Vulnetix found real bugs in our code, and we then audited the whole ecosystem for the same classes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;endpoints open without a key;&lt;/li&gt;
&lt;li&gt;checking a host name and then connecting to it;&lt;/li&gt;
&lt;li&gt;trusting forwarded client addresses.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;On November 5 we went through everything built for Multica looking for those classes, plus secrets reaching&lt;br&gt;
errors and logs, and allowlists that can be stepped around. The service, the scripts, the database function, the&lt;br&gt;
proxy and the ports were clean. Four things were not.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The order runner could reach every task in the workspace.&lt;/strong&gt; Its OpenCode rules allowed &lt;code&gt;multica issue get *&lt;/code&gt;,
and a Multica agent's token is bound to the workspace, not to its own task. Task numbers are sequential. Text a
buyer controls reaches the agent: the PR titles of the GitHub account they order. If that text talked the agent
into reading another task, the content could land in the summary. Multica posts that summary itself, outside the
tool calls a guard sees. A redirect (&lt;code&gt;&amp;gt; file&lt;/code&gt;) also passed the allowlist. The agent now runs a root-owned wrapper
that reads and closes only tasks assigned to it, with redirects, &lt;code&gt;$&lt;/code&gt; and backticks denied.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The guard could be renamed away.&lt;/strong&gt; Its files belonged to root, but the folder above them belonged to the agent
user, and Linux lets you rename a folder inside one you own. The guard now lives in a root-owned folder outside
the agent's home, and OpenCode finds it through &lt;code&gt;OPENCODE_CONFIG_DIR&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The memory-owner bug, one hub further.&lt;/strong&gt; When another hub routes a call to ours on its own key, our hub saw that
hub's account as the buyer, so all of its buyers would have shared one memory owner. Nobody had used that path.
Routed calls now name no owner.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Container addresses in our second hub's public manifest&lt;/strong&gt;, the same as the reviewer's finding. That fix is in
the code; this hub gets it with its next update.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Two things we learned about OpenCode 1.18.32 on the way:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Permissions.&lt;/strong&gt; It checks every command in a line on its own, including those inside &lt;code&gt;$(...)&lt;/code&gt;, backticks and
pipes, and the last matching rule wins. A redirected command is matched together with its redirect.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The config folder.&lt;/strong&gt; At start it writes a &lt;code&gt;.gitignore&lt;/code&gt; into every config folder. On a read-only mount it stops
with an error, so mount the guard's folder writable but owned by root.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We could not test the permission layer by asking the model to break it: DeepSeek refused all six bypass attempts on&lt;br&gt;
its own. We read OpenCode's source instead.&lt;/p&gt;

&lt;h2&gt;
  
  
  If you run Multica too
&lt;/h2&gt;

&lt;p&gt;Step by step, from the container to selling a long job, in English, Russian, Spanish, French and Chinese:&lt;br&gt;
&lt;a href="https://github.com/alexar76/aicom/tree/main/multica/manual" rel="noopener noreferrer"&gt;multica/manual&lt;/a&gt;. The Multica skill our agents use for&lt;br&gt;
buying and selling is &lt;a href="https://github.com/alexar76/aicom/blob/main/multica/community/aimarket/SKILL.md" rel="noopener noreferrer"&gt;SKILL.md&lt;/a&gt;.&lt;br&gt;
The full audit is in &lt;a href="https://github.com/alexar76/aicom/blob/main/artifacts/vulnetix-class-multica-2026-11-05/REPORT.md" rel="noopener noreferrer"&gt;REPORT.md&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;What we would tell another operator:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Keep facts in scripts and prose in models.&lt;/li&gt;
&lt;li&gt;Give each agent its own environment; the global config is shared by everyone on the daemon.&lt;/li&gt;
&lt;li&gt;Put the network allowlist below the prompt, and the guard below the model, in a folder the agent cannot touch.&lt;/li&gt;
&lt;li&gt;Allow an agent a wrapper you wrote, not a CLI with a wildcard.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This article was written with AI assistance from our agents' own logs and audit notes, and edited by a human.&lt;/p&gt;

&lt;p&gt;* We thank Vulnetix for helping us improve our products.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>opensource</category>
      <category>security</category>
    </item>
    <item>
      <title>Two MCP logs matched 23 tools. Five days later, every name had changed.</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Wed, 04 Nov 2026 08:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/two-mcp-logs-matched-23-tools-five-days-later-every-name-had-changed-1db0</link>
      <guid>https://dev.to/alexar76/two-mcp-logs-matched-23-tools-five-days-later-every-name-had-changed-1db0</guid>
      <description>&lt;p&gt;Compare the latest saved entries for one MCP server and you get zero tool names in common. Compare the observations from the same evening and all 23 tool fingerprints match.&lt;/p&gt;

&lt;p&gt;We maintain HISTOR, a transparency log for the tool definitions MCP servers advertise. We wanted to see whether an independent observer could corroborate one of our observations. We picked LEGION-tlog because it publishes a timestamp and a SHA-256 fingerprint for each tool. Those fingerprints let us compare individual definitions even when the logs use different formats for their records. &lt;a href="https://legion-api.com/tlog/docs/" rel="noopener noreferrer"&gt;LEGION developer documentation&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What the two logs saw
&lt;/h2&gt;

&lt;p&gt;The server was &lt;code&gt;com.aiapplyd/aiapplyd&lt;/code&gt;. We compared the tool name, description and input schema using LEGION’s canonicalization rule for per-tool fingerprints. All times below are UTC in 2026. &lt;a href="https://histor.modelmarket.dev/s/236eb5f0481cc99c" rel="noopener noreferrer"&gt;HISTOR server history&lt;/a&gt; · &lt;a href="https://tlog.legion-api.com/tlog/v0/servers/com.aiapplyd/aiapplyd" rel="noopener noreferrer"&gt;LEGION server record&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;HISTOR observation&lt;/th&gt;
&lt;th&gt;What the comparison showed&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;19 October, 21:37:07&lt;/td&gt;
&lt;td&gt;All 23 fingerprints matched the LEGION observation from 22:01:05, about 24 minutes later.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;24 October, 18:19:21&lt;/td&gt;
&lt;td&gt;All 23 names switched from &lt;code&gt;aiapplyd_&lt;/code&gt; to &lt;code&gt;getmattered_&lt;/code&gt;. The suffixes matched the earlier names. The latest entry in our archived LEGION response still showed the October 19 observation.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;31 October, 12:45:49&lt;/td&gt;
&lt;td&gt;HISTOR recorded a set of 40 tools. None of their names appeared in the archived LEGION record. This is the set in the November 2 HISTOR snapshot.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fun0yhfnrcu3wfwdzfe3p.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fun0yhfnrcu3wfwdzfe3p.webp" alt="Three HISTOR observations in UTC. On October 19 all 23 fingerprints match LEGION. On October 24 HISTOR sees 23 new names. On October 31 HISTOR sees 40 tools." width="800" height="336"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The October 24 observation is easy to misread if you only look at counts. Both sets contain 23 tools, but a client using the earlier names would have to account for the name changes. The matching suffixes give you a way to line up the names. Whether the tools behave the same would take a separate test.&lt;/p&gt;

&lt;p&gt;The evidence tells us when HISTOR first recorded the new set and which earlier version LEGION corroborated. It leaves open why the archived LEGION response had no later entry. In particular, the time of an observation and the time a server actually changed can differ.&lt;/p&gt;

&lt;h2&gt;
  
  
  Replay the check offline
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://modelmarket.dev/blog/media/histor-legion-offline-2026-11-03/replay.zip" rel="noopener noreferrer"&gt;Download the offline replay bundle&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;With the case bundle unpacked, run this from its directory using Node.js 22.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;node replay.mjs &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; replay-output.json
diff &lt;span class="nt"&gt;-u&lt;/span&gt; result.json replay-output.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;An empty diff means the replay reproduced the saved result. The script uses Node’s standard library and the bundled records, keys, signed tree heads (STHs) and inclusion proofs. It performs three checks.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;LEGION-tlog.&lt;/strong&gt; Verify the signing key against the pinned root key, verify the Ed25519 signatures on leaf 29030 and its STH, and recompute the path to the signed root of the 50,578-leaf Merkle tree.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;HISTOR.&lt;/strong&gt; Verify the STH and the three signed MTL/1 labels against HISTOR’s pinned &lt;code&gt;did:key&lt;/code&gt;. Verify their inclusion in the 1,564,894-leaf Merkle tree and recompute the tool set digest bound to each label.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Comparison.&lt;/strong&gt; Recompute the per-tool fingerprints from the authenticated HISTOR tool sets and compare them with the fingerprints in the LEGION record. The output lists names unique to either side and any shared names whose definitions differ.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The signed HISTOR label for the 40-tool set dates to October 31. The November 2 read time comes from snapshot metadata. The replay verifies inclusion under the bundled tree heads. Checking consistency across other tree heads or the behavior of the server would require separate evidence.&lt;/p&gt;

&lt;p&gt;For a tool-pinning workflow, the case suggests a concrete diagnostic. When two logs disagree, put the observation timestamps next to the fingerprints and look for an earlier matching tool set. That can turn a generic mismatch into a dated, inspectable change.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We maintain HISTOR. AI assistance was used for the analysis, the replay code, the illustrations and the writing.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>security</category>
      <category>cryptography</category>
      <category>opensource</category>
    </item>
    <item>
      <title>242,072 malware flags on npm and PyPI, and what the registries did about them</title>
      <dc:creator>Alex</dc:creator>
      <pubDate>Tue, 03 Nov 2026 14:00:00 +0000</pubDate>
      <link>https://dev.to/alexar76/242072-malware-flags-on-npm-and-pypi-and-what-the-registries-did-about-them-khn</link>
      <guid>https://dev.to/alexar76/242072-malware-flags-on-npm-and-pypi-and-what-the-registries-did-about-them-khn</guid>
      <description>&lt;p&gt;Security feeds publish lists of malicious packages, and dependency scanners turn those lists into&lt;br&gt;
red alerts in your CI. How often is a flag right? Nobody publishes that number per feed, so we&lt;br&gt;
started measuring it, with the method written down and committed before we looked at a single&lt;br&gt;
result. We also ran four open-source scanners on known malware and on a week of new uploads, to see&lt;br&gt;
what a flag costs at a registry's real volume. For completeness of the analysis, we also measured&lt;br&gt;
what each feed's records show and how each feed answers the people who dispute them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Our interest, first.&lt;/strong&gt; One of the feeds below listed nine of our own PyPI packages as malware; it has since reviewed all&lt;br&gt;
nine, withdrawing one and reclassifying eight as vulnerabilities. That is why we looked. They are counted like every other row, each feed gets its numbers a week before&lt;br&gt;
this post, and their answers are part of it. The last section shows how every tool treated our own&lt;br&gt;
packages, labelled for what it is: one interested publisher's case, not a measure of accuracy.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we measured
&lt;/h2&gt;

&lt;p&gt;Two open sources, npm and PyPI only:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;OpenSSF &lt;code&gt;malicious-packages&lt;/code&gt;: 234,710 records, each naming the tools that reported it. Nine tools
have attributable data: Amazon Inspector, ReversingLabs, Checkmarx, OSSF Package Analysis, Google
Open Source Security, GitHub's malware advisories, kam193, Oracle Macaron, and Vulnetix (7,362
malware records in its GCVE advisory database). Vulnetix points out that this database is not its
malware feed: it went public on 18 October and was still being loaded in bulk when we took our
snapshot on the 24th. Its malware feed proper is a set of STIX indicator bundles, which we did not
measure here.&lt;/li&gt;
&lt;li&gt;For a random sample of 7,483 records (up to 400 per tool, ecosystem and kind), we asked npm and
PyPI what they serve today. Metadata only: nothing was downloaded or run.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;"Registry acted" means the package is gone, replaced by npm's &lt;code&gt;0.0.1-security&lt;/code&gt; placeholder,&lt;br&gt;
quarantined, or the flagged versions are no longer served. That is a registry agreeing, not proof:&lt;br&gt;
registries act on these same feeds' reports. And a package still up is not proof of a false&lt;br&gt;
positive: registries are slow and miss things. This stage gives no false-positive rate. The next one&lt;br&gt;
does.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 1: most npm "malware" records are spam
&lt;/h2&gt;

&lt;p&gt;63.2% of OpenSSF's npm records describe the tea.xyz token-farming spam that flooded npm in late&lt;br&gt;
2025, almost all from one report by Amazon Inspector (140,729 records). Whether self-publishing&lt;br&gt;
reward spam is malware is a fair argument; Endor Labs called a similar set spam that steals&lt;br&gt;
nothing. We keep it apart. Registries acted on 56.0% of it (95% CI 51.1–60.8%).&lt;br&gt;
(A first count said 6.0%: npm also holds names as &lt;code&gt;0.0.1-security.N&lt;/code&gt;, and a held name can still list&lt;br&gt;
versions whose files are gone; our first check missed both. The same sample, checked again, is what&lt;br&gt;
this post reports.)&lt;/p&gt;

&lt;p&gt;On PyPI the picture is cleaner: 90.2% of records describe malicious code, 9.8% describe&lt;br&gt;
security-research probes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 2: how often registries acted, per tool
&lt;/h2&gt;

&lt;p&gt;For malicious-code flags 30 to 180 days old, registries acted on 87–100% of them for Amazon&lt;br&gt;
Inspector, GitHub, Google, kam193, OSSF Package Analysis and ReversingLabs. For Vulnetix's flags of&lt;br&gt;
the same age: 24.5% on npm and 11.0% on PyPI. (Intervals and sample sizes are in the full write-up.)&lt;/p&gt;

&lt;p&gt;We can't tell from this stage why, and there are several honest explanations:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The reporting path.&lt;/strong&gt; OpenSSF records reach npm and PyPI directly; some tools are PyPI trusted
reporters whose reports quarantine a project at once. A feed whose findings never reach the
registries' queues gets few takedowns whatever its accuracy. Vulnetix told us its findings don't
reach them: PyPI's security team has not answered its offer yet, and npm declined, since it runs
its own scanning.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A mix of kinds.&lt;/strong&gt; Vulnetix records carry no text to sort spam out, and registries act on spam
less often (56% against 92–100% for malicious code).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Different findings.&lt;/strong&gt; A feed can see malware the others miss.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;False positives.&lt;/strong&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Finding 3: how often feeds agree
&lt;/h2&gt;

&lt;p&gt;Google's npm flags are also flagged by another tool 95.2% of the time; ReversingLabs' about half the&lt;br&gt;
time. Low agreement is not a defect on its own: Checkmarx is alone on 99.5% of its PyPI packages,&lt;br&gt;
and PyPI removed every one we sampled. Vulnetix is alone on 97.7% of its PyPI packages and 90.0% of&lt;br&gt;
its npm ones, and registries left most of them up.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 4: one feed shows its work
&lt;/h2&gt;

&lt;p&gt;Vulnetix is the only feed here that publishes the rule and the matched line behind its records&lt;br&gt;
(28 of the 30 we drew).&lt;br&gt;
That deserves credit: it is the only reason anyone can check its reasoning.&lt;/p&gt;

&lt;p&gt;For completeness of the analysis, we also measured for every feed how transparent its records are&lt;br&gt;
and whether a verdict can be disputed, with the criteria fixed before we read the data (Finding 7).&lt;/p&gt;

&lt;p&gt;A quarter of its matches&lt;br&gt;
are indicator-list hits; the rest are text rules such as &lt;code&gt;/etc/passwd&lt;/code&gt; appearing in code (7.4%),&lt;br&gt;
&lt;code&gt;curl … | sh&lt;/code&gt; (7.2%), &lt;code&gt;python -c&lt;/code&gt; (5.8%), &lt;code&gt;~/.ssh&lt;/code&gt; (4.1%) and &lt;code&gt;exec(compile(…))&lt;/code&gt; (2.4%). These&lt;br&gt;
match anywhere in a package, tests and READMEs included. A USENIX Security 2026 paper found the same&lt;br&gt;
pattern in the open scanner GuardDog: most of its false positives were code-execution rules firing&lt;br&gt;
on legitimate idioms like &lt;code&gt;exec(compile(…))&lt;/code&gt;. We asked Vulnetix whether a record built on such a&lt;br&gt;
match means malware or a lead worth a look. It answered in the terms of the standards it publishes in:&lt;br&gt;
each record states its class (malware, weakness, exploit or vulnerability) and how it was reviewed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 5: among flags registries left up, how many are wrong
&lt;/h2&gt;

&lt;p&gt;Stage 2 took flags older than 30 days whose packages the registries still serve, a random sample of&lt;br&gt;
up to 50 per tool and ecosystem, and read them. Two model raters from different vendors, blind to&lt;br&gt;
which tool flagged a package, labelled each one against a codebook written beforehand ("malicious" =&lt;br&gt;
sends credentials, keys, files or environment out, runs a fetched payload, installs persistence,&lt;br&gt;
destroys data, or hides any of that). They were 29/30 right on held-out control packages and agreed&lt;br&gt;
with each other at κ = 0.715 on the sample; every disagreement was decided by a person reading the&lt;br&gt;
full code, and a blind re-read of 17 packages changed one label.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tool&lt;/th&gt;
&lt;th&gt;Benign among flags left up&lt;/th&gt;
&lt;th&gt;Without records the feed withdrew&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Vulnetix, npm&lt;/td&gt;
&lt;td&gt;47 of 50 (94.0%, 83.8–97.9%)&lt;/td&gt;
&lt;td&gt;same&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Vulnetix, PyPI&lt;/td&gt;
&lt;td&gt;50 of 50 (100%, 92.9–100%)&lt;/td&gt;
&lt;td&gt;same&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Amazon Inspector, PyPI&lt;/td&gt;
&lt;td&gt;32 of 44 (72.7%, 58.2–83.7%)&lt;/td&gt;
&lt;td&gt;16 of 25 (64.0%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Amazon Inspector, npm&lt;/td&gt;
&lt;td&gt;7 of 18 (38.9%)&lt;/td&gt;
&lt;td&gt;4 of 15 (26.7%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Google OSS security, npm&lt;/td&gt;
&lt;td&gt;1 of 13 (7.7%)&lt;/td&gt;
&lt;td&gt;same&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OpenSSF Package Analysis, npm&lt;/td&gt;
&lt;td&gt;0 of 16&lt;/td&gt;
&lt;td&gt;same&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;This is the error rate among flags registries did not act on, the flags most likely to be wrong; it&lt;br&gt;
is not a tool's overall false-positive rate. For OpenSSF's own analysis and Google, a flag that stayed&lt;br&gt;
up is almost always malware the registry missed. For Vulnetix, it is almost always a package that does&lt;br&gt;
nothing harmful: most of its records are matches on a domain or a string in a file. A day after&lt;br&gt;
our snapshot, Vulnetix withdrew 92% of its npm and PyPI records (Finding 7), and by 2 November 96 of&lt;br&gt;
the 100 we rated were withdrawn. Amazon&lt;br&gt;
Inspector's benign ones are mostly odd rather than harmful (a squatted name, a leaked API key), which&lt;br&gt;
a broader "suspicious" might still cover; we asked them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 6: scanners catch most known malware, and bury it in false alarms
&lt;/h2&gt;

&lt;p&gt;Feeds are fed by scanners, so stage 3 ran four open ones with their defaults, offline in a sandbox,&lt;br&gt;
300 s per package: GuardDog 3.2.0 (DataDog), malcontent 1.27.1 (Chainguard), OSSGadget 0.2.38&lt;br&gt;
&lt;code&gt;detect-backdoor&lt;/code&gt; (Microsoft), and a one-line baseline, "an install hook that makes a network call or&lt;br&gt;
starts a process". Three sets: 390 known malware samples from DataDog's public dataset (2025-2026),&lt;br&gt;
396 popular packages no feed lists, and 397 random uploads from 10-17 October that no feed listed a&lt;br&gt;
week later.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Scanner&lt;/th&gt;
&lt;th&gt;Known malware caught (npm / PyPI)&lt;/th&gt;
&lt;th&gt;False alarms per 1,000 recent uploads (npm / PyPI)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;GuardDog&lt;/td&gt;
&lt;td&gt;80.5% / 78.9%&lt;/td&gt;
&lt;td&gt;235 / 318&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;malcontent&lt;/td&gt;
&lt;td&gt;63.5% / 70.0%&lt;/td&gt;
&lt;td&gt;135 / 239&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OSSGadget&lt;/td&gt;
&lt;td&gt;86.0% / 90.5%&lt;/td&gt;
&lt;td&gt;828 / 990&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;One-line baseline&lt;/td&gt;
&lt;td&gt;57.5% / 56.8%&lt;/td&gt;
&lt;td&gt;10 / 66&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Benchmarks usually stop there, on a set that is half malware. A registry is not. In that October week&lt;br&gt;
the open feeds flagged 20 of the 28,833 PyPI projects that released, 0.07%. At that rate, on PyPI's&lt;br&gt;
~4,100 releasing projects a day:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Scanner&lt;/th&gt;
&lt;th&gt;Precision&lt;/th&gt;
&lt;th&gt;False alarms a day&lt;/th&gt;
&lt;th&gt;Real ones a day&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;One-line baseline&lt;/td&gt;
&lt;td&gt;0.59%&lt;/td&gt;
&lt;td&gt;~270&lt;/td&gt;
&lt;td&gt;~1.6&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;malcontent&lt;/td&gt;
&lt;td&gt;0.20%&lt;/td&gt;
&lt;td&gt;~980&lt;/td&gt;
&lt;td&gt;~2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GuardDog&lt;/td&gt;
&lt;td&gt;0.17%&lt;/td&gt;
&lt;td&gt;~1,310&lt;/td&gt;
&lt;td&gt;~2.3&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OSSGadget&lt;/td&gt;
&lt;td&gt;0.06%&lt;/td&gt;
&lt;td&gt;~4,080&lt;/td&gt;
&lt;td&gt;~2.6&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Hundreds of false alarms for each real one, whichever you pick. That is not a reason to drop them; it&lt;br&gt;
is why they are triage for a person and not a verdict, and why a scanner's hit needs a reader before it becomes a published verdict. npm is worse: a far larger volume and a lower&lt;br&gt;
base rate.&lt;/p&gt;

&lt;p&gt;What fires on ordinary new packages: GuardDog's rules for reading the environment, reading files,&lt;br&gt;
starting processes, downloading and running, which is what CLIs and agent tooling do all day;&lt;br&gt;
malcontent's URL, host:port and HTTP POST patterns; OSSGadget's "LOLBAS" and "data exfiltration"&lt;br&gt;
patterns, which match README headings and &lt;code&gt;return this;&lt;/code&gt;. We searched all 397 recent packages for&lt;br&gt;
the codebook's behaviours and read every hit; none was malicious.&lt;/p&gt;

&lt;p&gt;Two more things. On npm, recall drops for newer malware: malcontent catches 87% of samples from the&lt;br&gt;
first half of 2025 and 42% from the first half of 2026. And OSSGadget's tool package ships without its&lt;br&gt;
own rules, so out of the box it reports nothing for every package and does not say so; we added the&lt;br&gt;
rules of the same version. Caveat: the malware set is DataDog's and GuardDog is DataDog's scanner, so&lt;br&gt;
if DataDog finds malware partly with GuardDog, GuardDog's recall here is flattering. malcontent's&lt;br&gt;
author adds that every scanner has likely been trained or tested on this dataset, so recall on it is&lt;br&gt;
probably optimistic for malware a scanner has not seen; the repeat measurement adds samples published&lt;br&gt;
after 24 October.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finding 7: who shows their work, and who answers
&lt;/h2&gt;

&lt;p&gt;For completeness of the analysis, we measured what each record shows and whether its verdict can be&lt;br&gt;
disputed, with the criteria written down before we read the data.&lt;/p&gt;

&lt;p&gt;What a record shows, in 30 random records per tool:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tool&lt;/th&gt;
&lt;th&gt;Names what was found&lt;/th&gt;
&lt;th&gt;Says why, in its own words&lt;/th&gt;
&lt;th&gt;Says how it was found&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Vulnetix&lt;/td&gt;
&lt;td&gt;28&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Checkmarx&lt;/td&gt;
&lt;td&gt;26&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Amazon Inspector&lt;/td&gt;
&lt;td&gt;19&lt;/td&gt;
&lt;td&gt;19&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;kam193&lt;/td&gt;
&lt;td&gt;13 (29 with its linked analysis)&lt;/td&gt;
&lt;td&gt;29&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Oracle Macaron&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;OSSF Package Analysis&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Google&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ReversingLabs&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GitHub&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;No tool does all three. Vulnetix shows the matched line and the rule; on 29 of the 30 its reason is a standard sentence&lt;br&gt;
rather than one about the package. Amazon Inspector's best records are real analyses: files, hosts, what the code does.&lt;br&gt;
GitHub's text is advice ("should be considered fully compromised"), and the 39,857 records that name&lt;br&gt;
no tool say only that the package "was found to contain malicious code".&lt;/p&gt;

&lt;p&gt;How feeds answer, over 100 public disputes since January 2025:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;OpenSSF:&lt;/strong&gt; 44 disputes, 34 answered, median 0.7 days; 24 records withdrawn or corrected; 10
never answered, the oldest for ten months.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GitHub:&lt;/strong&gt; 48, 36 answered, median 1.2 days; 26 changed; 6 package owners sent to npm support,
where its malware advisories come from; 12 never answered.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vulnetix:&lt;/strong&gt; 8 since its tracker opened on 21 October, one of them ours. All answered within
three days, all changed. A week is not twenty-two months, so this compares nothing yet.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Then, on 25 October, Vulnetix ran an audit and withdrew 6,762 of the 7,362 npm and PyPI records we&lt;br&gt;
studied (92%), each withdrawal published with a statement that names the rule, the file and line,&lt;br&gt;
and why it is not malicious. It came the day after Vulnetix reviewed the records challenged in its&lt;br&gt;
tracker, ours among them, and found no malicious behaviour in any of those packages. We saw nothing&lt;br&gt;
on that scale from any other feed. Findings 2, 3 and 5&lt;br&gt;
describe the feed as it stood the day before; the repeat measurement counts these withdrawals as&lt;br&gt;
made before our letters. Vulnetix explains that its database was in its first week, when an automatic&lt;br&gt;
approval rule published static-rule matches before human review was wired in. Of the 600 records it&lt;br&gt;
kept, 577 had been approved by that rule; by 2 November, after our letters, it had withdrawn 52 more&lt;br&gt;
and reclassified 5 as vulnerabilities. It has since published a dispute process in which a person&lt;br&gt;
decides every case.&lt;/p&gt;

&lt;h2&gt;
  
  
  What every tool said about our own packages
&lt;/h2&gt;

&lt;p&gt;We publish 77 packages (52 on PyPI, 25 on npm). Here is what each tool says about them. One&lt;br&gt;
publisher is a case, not a sample; read it as what a maintainer runs into.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Eight of the nine feeds&lt;/strong&gt; (Amazon Inspector, ReversingLabs, Checkmarx, OSSF Package Analysis,
Google, GitHub, kam193, Oracle Macaron): nothing. None of the 77 is in their records.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vulnetix:&lt;/strong&gt; nine PyPI packages listed as malware, June to October. Every matched line we traced
is a test, an example or a constant: a test asserting our fetch tool &lt;em&gt;refuses&lt;/em&gt; &lt;code&gt;file:///etc/passwd&lt;/code&gt;,
&lt;code&gt;curl&lt;/code&gt; examples in READMEs, Circle's USDC contract address, the secp256k1 curve order, a cron line
in install instructions, &lt;code&gt;key_path: ~/.ssh/id_rsa&lt;/code&gt; in a configuration guide, and the list of persistence locations our own sandbox &lt;em&gt;watches&lt;/em&gt; when it runs other people's packages.
We disputed six in their repository on 21 October; three more were found while writing this.
Vulnetix reviewed all nine by hand within days: one record withdrawn, eight reclassified from
malware to vulnerabilities. And those eight were real: an MCP server that answered anyone when no
key was set, a probe that resolved a name twice so a second answer could point it inside our
network, sanctions data read incompletely, a permission that survived a runtime change, among
others. We fixed all eight and released the fixes; Vulnetix checked the published packages, set
the fixed versions on all eight records, and opened five new ones for flaws we described in our
own fixes. &lt;strong&gt;Thank you, Vulnetix&lt;/strong&gt;: the review was quick and done in the open, it
found real flaws that we have since fixed, and every step of it is public.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GuardDog 3.2.0&lt;/strong&gt; (open source, DataDog), which we ran ourselves with defaults: no verdict, by
design. It lists heuristics for a person to review. Its "threat" heuristics fired on 21 of 86 scans
(the latest version of all 77, plus the nine versions Vulnetix flagged, four of which it also
questioned): the Telegram API address our hub uses to notify its owner, &lt;code&gt;.env&lt;/code&gt; and environment
reads, WARDEN's own detection patterns for credential paths, &lt;code&gt;ffmpeg&lt;/code&gt; run with its output
silenced, a regular expression for Cyrillic. Each is code doing what it says; a reviewer clears it,
and asking for that review is exactly what GuardDog promises.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;WARDEN&lt;/strong&gt;, our own MCP firewall, so read this twice as interested: it reads the tool definitions a
model sees, not package code, and gives our MCP servers no blocking match. It does attach advisory
notes: 12 on our hub's 13 tools, mostly "this tool takes a credential parameter", which is true.
37.5% of all servers in the registry get advisory notes; 1.8% get a blocking match.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The pattern, if one case can show one, is that text rules matching anywhere in a package, tests and&lt;br&gt;
docs included, meet packages that &lt;em&gt;describe&lt;/em&gt; attacks: security tools, sandboxes, test suites.&lt;br&gt;
Finding 5 shows how common that is beyond us.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's next
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The same measurement again, two to three weeks after the feeds got their numbers:&lt;/strong&gt; the same rows checked
against the registries, what each feed withdrew or reclassified (before and after our letters,
reported apart, since a letter can cause a change), whether registries removed the packages we
read as malicious and kept the ones we read as benign, and a second week of recent uploads through
the same scanners. The plan is in the methodology, committed before the run. Vulnetix's audit of
25 October counts as a change made before the letters. At Vulnetix's suggestion it also measures
Vulnetix's STIX feed and runs its CLI's malware scan, set up the way Vulnetix advised.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The feeds' answers to our letters:&lt;/strong&gt; Vulnetix answered twice, and both answers are quoted in full
in the write-up; kam193 and malcontent's author answered too. Every later answer is added with a
date.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The method follows the ten pitfalls of Arp et al. (USENIX Security 2022), TESSERACT's handling of&lt;br&gt;
time, and the labelling practice of Guo et al. (ASE 2026). The full write-up in five languages, the&lt;br&gt;
methodology with its pitfall checklist, and the scripts and seeds to repeat every number:&lt;br&gt;
&lt;a href="https://github.com/alexar76/aicom/blob/main/docs/malware-feed-accuracy.md" rel="noopener noreferrer"&gt;the full write-up&lt;/a&gt; · &lt;a href="https://github.com/alexar76/aicom/blob/main/research/malware-feed-accuracy/METHODOLOGY.md" rel="noopener noreferrer"&gt;the methodology&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you run one of these feeds and think we got something wrong, tell us; corrections go into the&lt;br&gt;
write-up with a date.&lt;/p&gt;

&lt;p&gt;AI assistance was used for the analysis, the code and the writing.&lt;/p&gt;

</description>
      <category>security</category>
      <category>opensource</category>
      <category>npm</category>
      <category>python</category>
    </item>
  </channel>
</rss>
