<?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: Mag Solutions</title>
    <description>The latest articles on DEV Community by Mag Solutions (@mag_solutions_ai).</description>
    <link>https://dev.to/mag_solutions_ai</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%2F4117801%2F11fd5259-44d7-4dc3-bb74-82d5c8822a92.png</url>
      <title>DEV Community: Mag Solutions</title>
      <link>https://dev.to/mag_solutions_ai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/mag_solutions_ai"/>
    <language>en</language>
    <item>
      <title>Field report: 22344 real pull requests scanned, 7.7% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 16 Nov 2026 09:38:25 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-22344-real-pull-requests-scanned-77-flagged-46-of-our-own-mistakes-fixed-in-public-2506</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-22344-real-pull-requests-scanned-77-flagged-46-of-our-own-mistakes-fixed-in-public-2506</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;22344&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;7.7%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;19036&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;18890&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;46&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;90&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;46 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 22344 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-11-16 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
    <item>
      <title>A five-day-old fork of a real WhatsApp library, still live, that auto-follows a channel on every connection</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 16 Nov 2026 08:45:12 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/a-five-day-old-fork-of-a-real-whatsapp-library-still-live-that-auto-follows-a-channel-on-every-1423</link>
      <guid>https://dev.to/mag_solutions_ai/a-five-day-old-fork-of-a-real-whatsapp-library-still-live-that-auto-follows-a-channel-on-every-1423</guid>
      <description>&lt;p&gt;On 16 November 2026, GitHub's Advisory Database &lt;a href="https://github.com/advisories/GHSA-grg7-4c3w-pv5j" rel="noopener noreferrer"&gt;reviewed &lt;code&gt;@xzvbailss/baileys&lt;/code&gt;&lt;/a&gt;&lt;br&gt;
as malware. The package is a fork of &lt;a href="https://www.npmjs.com/package/baileys" rel="noopener noreferrer"&gt;Baileys&lt;/a&gt;, the WebSocket&lt;br&gt;
library most WhatsApp automation tools in Node.js are built on. The fork does what Baileys does —&lt;br&gt;
and then, without telling the user, subscribes their WhatsApp account to a channel the publisher&lt;br&gt;
picked.&lt;/p&gt;

&lt;p&gt;The advisory credits OpenSSF's malicious-packages project, record &lt;code&gt;MAL-2026-19946&lt;/code&gt;.&lt;/p&gt;
&lt;h2&gt;
  
  
  What the real package looks like next to the fork
&lt;/h2&gt;

&lt;p&gt;The comparison is the whole story. &lt;code&gt;baileys&lt;/code&gt; (no scope) has existed since September 2022, has gone&lt;br&gt;
through 64 published versions, is maintained under &lt;a href="https://github.com/WhiskeySockets/Baileys" rel="noopener noreferrer"&gt;&lt;code&gt;WhiskeySockets/Baileys&lt;/code&gt;&lt;/a&gt;&lt;br&gt;
on GitHub, and its &lt;code&gt;registry.npmjs.org&lt;/code&gt; record still lists that repository URL.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;@xzvbailss/baileys&lt;/code&gt; was created on 11 November 2026 — five days before the advisory. Its&lt;br&gt;
&lt;code&gt;package.json&lt;/code&gt; &lt;code&gt;repository&lt;/code&gt; field doesn't point to GitHub at all; it points to a Telegram handle,&lt;br&gt;
&lt;code&gt;https://t.me/JustRxVz&lt;/code&gt;. Its description is refreshingly candid for a malicious package: "hello I&lt;br&gt;
am Rz, I am a student trying to develop a WhatsApp library via baileys and this is my baileys&lt;br&gt;
xzvbailss." The same maintainer account, &lt;code&gt;justrxvz&lt;/code&gt;, also owns an unscoped package called&lt;br&gt;
&lt;code&gt;xzvbailss&lt;/code&gt; — cited in the advisory as a separate malware record, &lt;code&gt;MAL-2026-18702&lt;/code&gt;.&lt;/p&gt;
&lt;h2&gt;
  
  
  What it does, by version
&lt;/h2&gt;

&lt;p&gt;The advisory describes two behaviours, both triggered after a delay so they don't fire during a&lt;br&gt;
quick test install:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Versions 1.0.0 and 5.0.0&lt;/strong&gt; (both published 11 November, a minute apart): a hardcoded channel
ID triggers a &lt;code&gt;FOLLOW&lt;/code&gt; query 2.5 minutes after the socket connects, retried every 5 seconds.
Errors are swallowed silently.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Version 6.0.0&lt;/strong&gt; (published a few hours later the same day): the channel ID is now
base64-encoded and decoded at load. The follow runs on every &lt;code&gt;connection.update&lt;/code&gt; "open" event,
after a 90-second delay, with failures logged only at debug level — quieter, and harder to spot
by reading the diff.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The advisory lists only those three versions as affected. The registry shows six more releases&lt;br&gt;
after them — &lt;code&gt;6.0.1&lt;/code&gt; through &lt;code&gt;6.0.9&lt;/code&gt;, all published within a few hours of each other on 13&lt;br&gt;
November, two days after the advisory's sample and three days before it was reviewed.&lt;/p&gt;
&lt;h2&gt;
  
  
  We checked the version that's live right now
&lt;/h2&gt;

&lt;p&gt;So we pulled the tarball for &lt;code&gt;6.0.9&lt;/code&gt;, the version &lt;code&gt;npm install @xzvbailss/baileys&lt;/code&gt; would install&lt;br&gt;
today, and looked for the behaviour ourselves. It's in &lt;code&gt;lib/Socket/newsletter.js&lt;/code&gt;:&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="c1"&gt;// XzV Newsletter Auto-Follow: follow one configured newsletter after connection.&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;newsletterAutoFollowId&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;config&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;newsletterAutoFollowId&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="s1"&gt;'120363430764800148@newsletter'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;AUTO_FOLLOW_DELAY&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;10_000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;AUTO_FOLLOW_RETRY_DELAY&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;20_000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;AUTO_FOLLOW_MAX_ATTEMPTS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No base64 this time — the comment names exactly what the code does. A timer fires 10 seconds after&lt;br&gt;
the socket opens, calls the same &lt;code&gt;FOLLOW&lt;/code&gt; query ID the advisory describes, retries up to three&lt;br&gt;
times 20 seconds apart, and skips itself quietly if the configured ID isn't a well-formed&lt;br&gt;
newsletter JID. The hardcoded channel differs from the one the advisory quotes for &lt;code&gt;6.0.0&lt;/code&gt;, which&lt;br&gt;
tells us the operator changed the target at least once across the six patch releases — not that the&lt;br&gt;
behaviour ever went away.&lt;/p&gt;

&lt;p&gt;In short: GitHub's "affected versions" field is accurate for the versions it reviewed, but the&lt;br&gt;
package kept shipping the same feature after that review window, and nothing currently stops&lt;br&gt;
&lt;code&gt;npm install @xzvbailss/baileys&lt;/code&gt; from pulling it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The family this sits in
&lt;/h2&gt;

&lt;p&gt;The advisory ties &lt;code&gt;@xzvbailss/baileys&lt;/code&gt; to two other records: &lt;code&gt;MAL-2026-17342&lt;/code&gt;, a "PhantomSub"&lt;br&gt;
pattern it says matches the same behaviour, and &lt;code&gt;MAL-2026-18873&lt;/code&gt;, a different package,&lt;br&gt;
&lt;code&gt;@poucode/baileys&lt;/code&gt;, that the advisory says shares a byte-identical copy of the same &lt;code&gt;luxu.js&lt;/code&gt; file.&lt;br&gt;
We checked: &lt;code&gt;@poucode/baileys&lt;/code&gt; exists, was created on 11 October 2026 — a month before the&lt;br&gt;
&lt;code&gt;xzvbailss&lt;/code&gt; fork — and is also still resolvable, currently at version &lt;code&gt;2.0.1&lt;/code&gt;. We didn't go further&lt;br&gt;
and compare the two packages' files ourselves; that claim is the advisory's, not something we&lt;br&gt;
re-verified byte for byte.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this fits on the pull request
&lt;/h2&gt;

&lt;p&gt;There's no typo here and no install script. &lt;code&gt;@xzvbailss/baileys&lt;/code&gt; isn't trying to look like&lt;br&gt;
&lt;code&gt;baileys&lt;/code&gt; by misspelling it — it's an entirely different, scoped package name, and nobody adds it&lt;br&gt;
to a project by accident. The signal a check can act on isn't the name; it's that the package is&lt;br&gt;
five days old with no history, forked from something 64 releases and four years deep, pointing at a&lt;br&gt;
Telegram handle instead of a repository. That's exactly the "newly published, low-adoption&lt;br&gt;
dependency" pattern MagAudit's check is built to flag on the pull request that adds the dependency&lt;br&gt;
line — whether that stops the merge depends on whether the team has made the check required in&lt;br&gt;
their own branch protection settings, the same as any other check on the PR. A general-purpose AI&lt;br&gt;
reviewer reading &lt;code&gt;"@xzvbailss/baileys": "^6.0.9"&lt;/code&gt; in a diff has nothing in the line itself to go on;&lt;br&gt;
the age and the fork relationship only show up by querying the registry, which is what a dedicated&lt;br&gt;
check is there to do.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this doesn't tell you
&lt;/h2&gt;

&lt;p&gt;The advisory doesn't say what the followed channel publishes, how many accounts the auto-follow has&lt;br&gt;
actually reached, or whether &lt;code&gt;justrxvz&lt;/code&gt; controls the channel directly or is running this for someone&lt;br&gt;
else. We don't know why the operator moved from a silent 2.5-minute delay to a candidly-commented&lt;br&gt;
10-second one — confidence that reviewers wouldn't catch it, indifference, or simply sloppy&lt;br&gt;
iteration. We didn't check versions &lt;code&gt;5.0.0&lt;/code&gt; through &lt;code&gt;6.0.8&lt;/code&gt; individually; our own verification&lt;br&gt;
covers the advisory's description of &lt;code&gt;1.0.0&lt;/code&gt;/&lt;code&gt;5.0.0&lt;/code&gt;/&lt;code&gt;6.0.0&lt;/code&gt; and our own read of &lt;code&gt;6.0.9&lt;/code&gt;, nothing in&lt;br&gt;
between. And we took the advisory's word for the shared &lt;code&gt;luxu.js&lt;/code&gt; file with &lt;code&gt;@poucode/baileys&lt;/code&gt;&lt;br&gt;
rather than diffing the two packages ourselves.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>javascript</category>
      <category>ai</category>
    </item>
    <item>
      <title>Field report: 20359 real pull requests scanned, 6.5% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 09 Nov 2026 09:39:15 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-20359-real-pull-requests-scanned-65-flagged-46-of-our-own-mistakes-fixed-in-public-26bo</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-20359-real-pull-requests-scanned-65-flagged-46-of-our-own-mistakes-fixed-in-public-26bo</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;20359&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6.5%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;15988&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;16105&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;82&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;40 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 20359 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-11-09 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
    <item>
      <title>Field report: 17381 real pull requests scanned, 10% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 02 Nov 2026 09:34:03 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-17381-real-pull-requests-scanned-10-flagged-46-of-our-own-mistakes-fixed-in-public-4ob6</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-17381-real-pull-requests-scanned-10-flagged-46-of-our-own-mistakes-fixed-in-public-4ob6</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;17381&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;10%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;14313&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;14483&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;33&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;70&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;33 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 17381 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-11-02 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
    <item>
      <title>Drop the scope, keep the name: six packages squatting Axelar, Solana and OpenZeppelin SDKs in 48 hours</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 02 Nov 2026 08:44:33 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/drop-the-scope-keep-the-name-six-packages-squatting-axelar-solana-and-openzeppelin-sdks-in-48-1af6</link>
      <guid>https://dev.to/mag_solutions_ai/drop-the-scope-keep-the-name-six-packages-squatting-axelar-solana-and-openzeppelin-sdks-in-48-1af6</guid>
      <description>&lt;p&gt;Between 01:24 UTC on 1 November 2026 and 02:33 UTC on 2 November 2026, two npm accounts —&lt;br&gt;
&lt;code&gt;cutekitten05&lt;/code&gt; and &lt;code&gt;bigdevin55&lt;/code&gt; — published a cluster of packages built around one idea: take a&lt;br&gt;
real, scoped SDK name and republish it without the scope. &lt;a href="https://github.com/advisories/GHSA-69w5-x7cm-v78h" rel="noopener noreferrer"&gt;&lt;code&gt;axelarjs-sdk&lt;/code&gt;&lt;/a&gt;&lt;br&gt;
copies &lt;code&gt;@axelar-network/axelarjs-sdk&lt;/code&gt;. &lt;a href="https://github.com/advisories/GHSA-r89f-jjgp-3q97" rel="noopener noreferrer"&gt;&lt;code&gt;wallet-adapter-base&lt;/code&gt;&lt;/a&gt;&lt;br&gt;
copies &lt;code&gt;@solana/wallet-adapter-base&lt;/code&gt;. &lt;a href="https://github.com/advisories/GHSA-j62h-2r3r-ggh9" rel="noopener noreferrer"&gt;&lt;code&gt;sol-wallet-adapter&lt;/code&gt;&lt;/a&gt;&lt;br&gt;
copies &lt;code&gt;@project-serum/sol-wallet-adapter&lt;/code&gt;. &lt;a href="https://github.com/advisories/GHSA-v7f8-27g9-qx88" rel="noopener noreferrer"&gt;&lt;code&gt;contracts-upgradeable&lt;/code&gt;&lt;/a&gt;&lt;br&gt;
copies &lt;code&gt;@openzeppelin/contracts-upgradeable&lt;/code&gt;. GitHub's Advisory Database reviewed all four as&lt;br&gt;
malware on 2 November, crediting the OpenSSF malicious-packages project (records MAL-2026-18456,&lt;br&gt;
18459, 18452 and 18451).&lt;/p&gt;

&lt;p&gt;The same accounts also published two packages under their own names —&lt;br&gt;
&lt;a href="https://github.com/advisories/GHSA-pw4j-9v86-8c3g" rel="noopener noreferrer"&gt;&lt;code&gt;cutekitten-test&lt;/code&gt;&lt;/a&gt;, whose README claims it is&lt;br&gt;
"a security test package for a school project," and &lt;a href="https://github.com/advisories/GHSA-j2hf-rv9p-8jvr" rel="noopener noreferrer"&gt;&lt;code&gt;bigbob22-test&lt;/code&gt;&lt;/a&gt;,&lt;br&gt;
with an empty README. Every one of the six ships the identical script.&lt;/p&gt;
&lt;h2&gt;
  
  
  One script, six names
&lt;/h2&gt;

&lt;p&gt;Each advisory describes the same &lt;code&gt;CWE-506&lt;/code&gt; behaviour: the package "has no functional code." A&lt;br&gt;
&lt;code&gt;preinstall&lt;/code&gt; script runs &lt;code&gt;node preinstall.js&lt;/code&gt; during &lt;code&gt;npm install&lt;/code&gt;, and the script sends the&lt;br&gt;
machine's hostname, OS username and local time to a Discord webhook the advisories label "PC&lt;br&gt;
Notifier." None of the six does anything else — no real SDK, no stub exports, nothing to make a&lt;br&gt;
build keep working. If a project actually tries to import &lt;code&gt;axelarjs-sdk&lt;/code&gt; expecting the Axelar&lt;br&gt;
client, it breaks immediately; the only "feature" that reliably works is the beacon firing on&lt;br&gt;
install.&lt;/p&gt;

&lt;p&gt;That's a smaller ambition than campaigns that ship a working no-op shim to avoid crashing the&lt;br&gt;
build. This one doesn't bother. Whoever is behind it isn't trying to stay resident in a dependency&lt;br&gt;
tree — they just want the install to happen once.&lt;/p&gt;
&lt;h2&gt;
  
  
  Checking the registry directly
&lt;/h2&gt;

&lt;p&gt;The advisories give affected versions; the registry's own &lt;code&gt;time&lt;/code&gt; and &lt;code&gt;dist-tags&lt;/code&gt; fields show what&lt;br&gt;
state each name is in right now:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/wallet-adapter-base | python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"
import json, sys
d = json.load(sys.stdin)
print(d['time'])
print(d['dist-tags'])
"&lt;/span&gt;
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'created'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-11-01T21:32:11.699Z'&lt;/span&gt;, &lt;span class="s1"&gt;'modified'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-11-01T21:33:07.093Z'&lt;/span&gt;,
 &lt;span class="s1"&gt;'0.0.0-stage'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-11-01T21:32:11.699Z'&lt;/span&gt;, &lt;span class="s1"&gt;'1.0.0'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-11-01T21:33:06.920Z'&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'latest'&lt;/span&gt;: &lt;span class="s1"&gt;'1.0.0'&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;wallet-adapter-base&lt;/code&gt; went from a stub to its only real release, &lt;code&gt;1.0.0&lt;/code&gt; — the version the advisory&lt;br&gt;
names as malicious — in under a minute, and &lt;code&gt;latest&lt;/code&gt; still points there. Running the same query&lt;br&gt;
against &lt;code&gt;sol-wallet-adapter&lt;/code&gt; and &lt;code&gt;contracts-upgradeable&lt;/code&gt; shows the same thing: &lt;code&gt;latest&lt;/code&gt; resolves to&lt;br&gt;
&lt;code&gt;1.0.7&lt;/code&gt; and &lt;code&gt;1.0.0&lt;/code&gt; respectively, both the versions the advisories flag. &lt;code&gt;npm install&lt;br&gt;
wallet-adapter-base&lt;/code&gt; today still pulls the beacon.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;axelarjs-sdk&lt;/code&gt; is the exception. Its registry record now shows a single version, &lt;code&gt;0.0.0-stage&lt;/code&gt;,&lt;br&gt;
with a &lt;code&gt;created&lt;/code&gt; timestamp of &lt;code&gt;2026-11-02T02:30:26Z&lt;/code&gt; — after the account's original &lt;code&gt;1.0.1&lt;/code&gt; release&lt;br&gt;
the advisory describes. &lt;code&gt;cutekitten-test&lt;/code&gt; and &lt;code&gt;bigbob22-test&lt;/code&gt; show an &lt;code&gt;unpublished&lt;/code&gt; block in their&lt;br&gt;
&lt;code&gt;time&lt;/code&gt; field instead of a current version at all. That's npm's usual pattern after a takedown: pull&lt;br&gt;
the malicious release and leave an empty placeholder, or remove the name's history outright, rather&lt;br&gt;
than simply flagging it. It also means "reviewed as malware" and "no longer installable" are two&lt;br&gt;
different facts, and as of this writing only two of the six packages have both.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this fits on the pull request
&lt;/h2&gt;

&lt;p&gt;Nothing here depends on a misspelling. &lt;code&gt;axelarjs-sdk&lt;/code&gt; isn't a typo of &lt;code&gt;@axelar-network/axelarjs-sdk&lt;/code&gt;&lt;br&gt;
— it's the same name with the scope removed, which is exactly the string a developer might type&lt;br&gt;
from memory, or that an AI assistant might generate if it drops the scope while filling in an&lt;br&gt;
import. The same is true for &lt;code&gt;wallet-adapter-base&lt;/code&gt; against &lt;code&gt;@solana/wallet-adapter-base&lt;/code&gt;. Neither&lt;br&gt;
requires the attacker to guess anything; the real package's own name is the bait.&lt;/p&gt;

&lt;p&gt;On a pull request that adds one of these as a dependency, that's two independent signals a check&lt;br&gt;
can act on without knowing anything about Discord webhooks: a package name a few hours or days old&lt;br&gt;
with no history, and a &lt;code&gt;preinstall&lt;/code&gt; script that runs code before anything else in the project does.&lt;br&gt;
That's what MagAudit marks as critical on the lockfile line — whether it would have stopped the&lt;br&gt;
merge depends on whether the team has made the check required in their own branch protection&lt;br&gt;
settings. A general-purpose AI reviewer reading the diff has little reason to flag it: &lt;code&gt;"wallet-adapter-base": "^1.0.0"&lt;/code&gt;&lt;br&gt;
is syntactically ordinary, and nothing about the line itself resolves against the registry.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this doesn't tell you
&lt;/h2&gt;

&lt;p&gt;The advisories don't say who received the beacons, whether the Discord webhook had an audience&lt;br&gt;
beyond its operator, or why a six-package campaign bothers squatting real SDK names at all when&lt;br&gt;
the payload collects nothing more sensitive than a hostname and username — that's a long way from&lt;br&gt;
the credential or session-token theft seen in other campaigns we've written about. We also don't&lt;br&gt;
know whether &lt;code&gt;cutekitten05&lt;/code&gt; and &lt;code&gt;bigdevin55&lt;/code&gt; are the same person, or whether "security test package&lt;br&gt;
for a school project" is a genuine excuse, a decoy, or an attempt to pre-empt exactly the kind of&lt;br&gt;
writeup this is. We opened the four squat advisories and the two self-named ones directly; we&lt;br&gt;
haven't independently confirmed the OpenSSF OSV records beyond what each GitHub advisory cites, and&lt;br&gt;
we didn't attempt to resolve or contact the webhook endpoint.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>javascript</category>
      <category>ai</category>
    </item>
    <item>
      <title>Field report: 14898 real pull requests scanned, 6.4% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 26 Oct 2026 09:33:43 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-14898-real-pull-requests-scanned-64-flagged-46-of-our-own-mistakes-fixed-in-public-jhf</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-14898-real-pull-requests-scanned-64-flagged-46-of-our-own-mistakes-fixed-in-public-jhf</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;14898&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6.4%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;12740&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;12885&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;25&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;60&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;25 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 14898 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-10-26 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
    <item>
      <title>One npm account, 14 packages, four scopes: a dependency-confusion campaign built to look like someone's internal SDK</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 26 Oct 2026 08:40:55 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/one-npm-account-14-packages-four-scopes-a-dependency-confusion-campaign-built-to-look-like-44ne</link>
      <guid>https://dev.to/mag_solutions_ai/one-npm-account-14-packages-four-scopes-a-dependency-confusion-campaign-built-to-look-like-44ne</guid>
      <description>&lt;p&gt;Between 21:23 UTC on 23 October 2026 and 14:28 UTC on 25 October 2026, an npm account called&lt;br&gt;
&lt;code&gt;wiseke&lt;/code&gt; published 14 packages across four scopes: &lt;code&gt;@greenlake&lt;/code&gt;, &lt;code&gt;@sxm-web&lt;/code&gt;, &lt;code&gt;@drone-js-ext&lt;/code&gt; and&lt;br&gt;
&lt;code&gt;@ppa-web-components&lt;/code&gt;. On 26 October, GitHub's Advisory Database reviewed all of them as malware,&lt;br&gt;
including &lt;a href="https://github.com/advisories/GHSA-pfqf-544g-rvp9" rel="noopener noreferrer"&gt;&lt;code&gt;@greenlake/auth&lt;/code&gt;&lt;/a&gt;,&lt;br&gt;
&lt;a href="https://github.com/advisories/GHSA-pmfr-35rg-27xm" rel="noopener noreferrer"&gt;&lt;code&gt;@greenlake/context-pane&lt;/code&gt;&lt;/a&gt;,&lt;br&gt;
&lt;a href="https://github.com/advisories/GHSA-968q-w3j5-w6rq" rel="noopener noreferrer"&gt;&lt;code&gt;@greenlake/mfe-login-url&lt;/code&gt;&lt;/a&gt; and&lt;br&gt;
&lt;a href="https://github.com/advisories/GHSA-j5p5-hcm3-6gx9" rel="noopener noreferrer"&gt;&lt;code&gt;@sxm-web/client-sdk-core&lt;/code&gt;&lt;/a&gt;. Each advisory&lt;br&gt;
describes the same package, published under a different scoped name.&lt;/p&gt;

&lt;p&gt;Scoped names like &lt;code&gt;@greenlake/auth&lt;/code&gt; or &lt;code&gt;@greenlake/mfe-login-url&lt;/code&gt; read like the internal packages&lt;br&gt;
of a real micro-frontend codebase — an auth module, a context pane, a login-URL helper for a&lt;br&gt;
federated UI. That's the point of dependency confusion: if an organisation has an internal package&lt;br&gt;
under a given scope and a build or CI step ever resolves that scope against the public registry&lt;br&gt;
instead of (or before) the private one, npm installs whichever version is public — in this case,&lt;br&gt;
one the attacker controls.&lt;/p&gt;
&lt;h2&gt;
  
  
  What the advisories say it does
&lt;/h2&gt;

&lt;p&gt;All four advisories describe identical behaviour, classified &lt;code&gt;CWE-506&lt;/code&gt;, Embedded Malicious Code:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The package calls itself a "Compatibility shim" and its README describes an "authorized
dependency-confusion test" — a label the package itself supplies, not something GitHub verified.&lt;/li&gt;
&lt;li&gt;Each affected version ships a &lt;code&gt;postinstall&lt;/code&gt; script running &lt;code&gt;node beacon.cjs&lt;/code&gt;, and &lt;code&gt;index.js&lt;/code&gt;
runs the same code again on import.&lt;/li&gt;
&lt;li&gt;The code sends the machine's hostname, install path and current working directory over &lt;strong&gt;plain
HTTP&lt;/strong&gt; to &lt;code&gt;185.158.107.175:8787&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;index.js&lt;/code&gt; exports a Proxy that returns no-op functions for everything, so a build that imports
the package keeps compiling instead of throwing on a missing export — nothing crashes to draw
attention to it.&lt;/li&gt;
&lt;li&gt;GitHub's advisories note that this &lt;code&gt;index.js&lt;/code&gt; is byte-identical to &lt;code&gt;ph-common&lt;/code&gt; (MAL-2026-1809),
which beacons to the same IP and port — the same toolkit reused across more than one campaign.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Both &lt;code&gt;@sxm-web/client-sdk-core&lt;/code&gt; and &lt;code&gt;@greenlake/context-pane&lt;/code&gt; list a version string of &lt;code&gt;13.3.7&lt;/code&gt; —&lt;br&gt;
not a plausible result of normal semantic versioning, but a number chosen to look like a mature,&lt;br&gt;
long-maintained package rather than something published an hour earlier.&lt;/p&gt;
&lt;h2&gt;
  
  
  Checking the timeline against the registry directly
&lt;/h2&gt;

&lt;p&gt;The advisories give a publication window; the registry's own &lt;code&gt;time&lt;/code&gt; field confirms how fast the&lt;br&gt;
first package in the batch went from created to malicious:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/@greenlake/auth | python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"
import json, sys
print(json.load(sys.stdin)['time'])
"&lt;/span&gt;
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'created'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-23T21:23:50.663Z'&lt;/span&gt;, &lt;span class="s1"&gt;'modified'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-23T21:26:05.976Z'&lt;/span&gt;,
 &lt;span class="s1"&gt;'0.0.0-stage'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-23T21:23:50.663Z'&lt;/span&gt;, &lt;span class="s1"&gt;'0.1.0'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-23T21:25:28.491Z'&lt;/span&gt;,
 &lt;span class="s1"&gt;'3.0.0'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-23T21:26:05.660Z'&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;@greenlake/auth&lt;/code&gt; went from its first stub version to the malicious &lt;code&gt;3.0.0&lt;/code&gt; release in under three&lt;br&gt;
minutes. The package has existed, in total, for less time than it would take to read its own&lt;br&gt;
README.&lt;/p&gt;

&lt;p&gt;Querying the registry today shows npm has already intervened on some of the names:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/@greenlake/auth | python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"
import json, sys
print(json.load(sys.stdin)['dist-tags'])
"&lt;/span&gt;
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'latest'&lt;/span&gt;: &lt;span class="s1"&gt;'0.0.0-stage'&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;latest&lt;/code&gt; now points back to the harmless stub version that was published first, which is how npm&lt;br&gt;
quietly neutralises a scope without deleting the historical record the advisory refers to.&lt;br&gt;
&lt;code&gt;@sxm-web/client-sdk-core&lt;/code&gt; still resolves &lt;code&gt;latest&lt;/code&gt; to &lt;code&gt;13.3.7&lt;/code&gt; — the malicious version — at the&lt;br&gt;
time of writing, which is a reminder that an advisory existing is not the same as a registry&lt;br&gt;
already being clean.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this fits on the pull request
&lt;/h2&gt;

&lt;p&gt;Nothing about this campaign depends on a typo. The names aren't misspellings of real packages;&lt;br&gt;
they're scopes an attacker is betting some organisation uses internally, published with inflated&lt;br&gt;
version numbers to look established rather than new. That makes two of the signals a pull request&lt;br&gt;
check can act on: a dependency that didn't exist a week ago, and a lockfile addition with no&lt;br&gt;
matching entry in whatever private registry the scope is supposed to belong to. A &lt;code&gt;postinstall&lt;/code&gt;&lt;br&gt;
script that runs on every install is a third, independent of whether the name looks suspicious at&lt;br&gt;
all.&lt;/p&gt;

&lt;p&gt;On a pull request that adds one of these as a dependency, that's what MagAudit marks as critical,&lt;br&gt;
with the file and line of the lockfile change — whether that would have stopped the merge depends&lt;br&gt;
on whether the team made the check required in their own branch protection settings. It's also&lt;br&gt;
exactly the kind of thing a general-purpose AI code reviewer, reading the diff for logic and style,&lt;br&gt;
has no particular reason to flag: the new line in &lt;code&gt;package.json&lt;/code&gt; is syntactically fine, and nothing&lt;br&gt;
about &lt;code&gt;"@greenlake/auth": "^3.0.0"&lt;/code&gt; looks wrong without checking the registry against it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this doesn't tell you
&lt;/h2&gt;

&lt;p&gt;The advisories don't say who, if anyone, actually installed these packages, so we don't know&lt;br&gt;
whether the beacon ever fired against a real organisation's machine, or how many of the other ten&lt;br&gt;
packages in the batch we haven't individually checked follow the identical pattern — we're&lt;br&gt;
confident they do, based on GitHub's description of all 14 as one campaign, but we only opened the&lt;br&gt;
four advisories above. We also don't know who operates &lt;code&gt;185.158.107.175&lt;/code&gt;, whether &lt;code&gt;wiseke&lt;/code&gt; is a&lt;br&gt;
single person or a shared tooling account, or whether the real targets are the four named scopes or&lt;br&gt;
unrelated organisations that happen to use similarly named internal packages. The "authorized&lt;br&gt;
dependency-confusion test" label in the README is the account's own claim; nothing in the&lt;br&gt;
advisories or the registry confirms or denies it, and a real authorised test wouldn't normally&lt;br&gt;
need to hide its network calls behind a no-op Proxy.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>javascript</category>
      <category>ai</category>
    </item>
    <item>
      <title>Field report: 12656 real pull requests scanned, 9.6% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 19 Oct 2026 09:38:47 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-12656-real-pull-requests-scanned-96-flagged-46-of-our-own-mistakes-fixed-in-public-4412</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-12656-real-pull-requests-scanned-96-flagged-46-of-our-own-mistakes-fixed-in-public-4412</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;12656&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;9.6%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;11266&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;11494&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;51&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;20 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 12656 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-10-19 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
    <item>
      <title>items-validator and item-validator: three weeks old and six years dormant, flagged as malware the same morning</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 19 Oct 2026 08:46:18 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/items-validator-and-item-validator-three-weeks-old-and-six-years-dormant-flagged-as-malware-the-p66</link>
      <guid>https://dev.to/mag_solutions_ai/items-validator-and-item-validator-three-weeks-old-and-six-years-dormant-flagged-as-malware-the-p66</guid>
      <description>&lt;p&gt;On the morning of 19 October 2026, GitHub's Advisory Database reviewed two npm packages as&lt;br&gt;
malware: &lt;a href="https://github.com/advisories/GHSA-rpgc-c7g7-g8fh" rel="noopener noreferrer"&gt;&lt;code&gt;items-validator&lt;/code&gt;&lt;/a&gt; and&lt;br&gt;
&lt;a href="https://github.com/advisories/GHSA-ff48-h3v7-m43r" rel="noopener noreferrer"&gt;&lt;code&gt;item-validator&lt;/code&gt;&lt;/a&gt; — names one letter apart,&lt;br&gt;
in the same batch of advisories. Looked at side by side on the registry, they turn out to be two&lt;br&gt;
different supply-chain patterns wearing the same name.&lt;/p&gt;
&lt;h2&gt;
  
  
  What the advisories actually say
&lt;/h2&gt;

&lt;p&gt;Both are short, and both say the same thing in the same words: "Any computer that has this&lt;br&gt;
package installed or running should be considered fully compromised," with a recommendation to&lt;br&gt;
rotate every secret on the affected machine from a different computer and remove the package,&lt;br&gt;
while warning that removal alone may not clean up whatever it already did. Both are classified&lt;br&gt;
&lt;code&gt;CWE-506&lt;/code&gt;, "Embedded Malicious Code." Neither advisory describes the actual payload, names a CVE,&lt;br&gt;
or links to the code — this is the standard malware-advisory boilerplate GitHub publishes when a&lt;br&gt;
package is pulled, not a writeup of what it does. We're not going to guess at the mechanism from&lt;br&gt;
two sentences; what follows is what we could check ourselves, directly against the registry.&lt;/p&gt;
&lt;h2&gt;
  
  
  One package is three weeks old. The other is six years dormant.
&lt;/h2&gt;

&lt;p&gt;The registry's own &lt;code&gt;time&lt;/code&gt; field — the publish timestamp of every version a package has ever had —&lt;br&gt;
is public and doesn't get rewritten when npm pulls a package for malware, even though the actual&lt;br&gt;
version contents do:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/items-validator | python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"import json,sys; print(json.load(sys.stdin)['time'])"&lt;/span&gt;
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'1.0.4'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-09-28T08:37:45.576Z'&lt;/span&gt;, &lt;span class="s1"&gt;'1.0.5'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-09-28T12:00:19.110Z'&lt;/span&gt;,
 &lt;span class="s1"&gt;'0.0.0-stage'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-19T06:24:58.173Z'&lt;/span&gt;, &lt;span class="s1"&gt;'0.0.1-security'&lt;/span&gt;: &lt;span class="s1"&gt;'2026-10-19T06:25:55.248Z'&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;items-validator&lt;/code&gt; had exactly two real versions, both published 28 September — three weeks before&lt;br&gt;
the advisory. That's the textbook shape of a newly-published, low-adoption dependency: a handful&lt;br&gt;
of versions, no track record, added to someone's &lt;code&gt;package.json&lt;/code&gt; while it was still fresh.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;item-validator&lt;/code&gt; (singular) tells a completely different story:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/item-validator | python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"import json,sys; d=json.load(sys.stdin)['time']; print(min(d.values()), max(v for k,v in d.items() if not k.startswith('0.0')))"&lt;/span&gt;
2019-12-30T14:30:58.588Z 2020-02-20T21:05:26.313Z
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That package shipped more than 70 versions between December 2019 and February 2020, then nothing&lt;br&gt;
— no release, no advisory, no activity of any kind — for over six years. Then, this morning, it&lt;br&gt;
was quarantined for the same reason as its three-week-old near-namesake. Both registry entries now&lt;br&gt;
show &lt;code&gt;"maintainers": [{"name": "npm", "email": "npm@npmjs.com"}]&lt;/code&gt; and&lt;br&gt;
&lt;code&gt;"description": "security holding package"&lt;/code&gt; — npm's standard placeholder once it pulls a package&lt;br&gt;
for malware and takes over the name itself.&lt;/p&gt;

&lt;p&gt;A package dormant for six years and then suddenly shipping new versions is a different failure&lt;br&gt;
mode from a brand-new package with no history: it usually means an attacker got control of an&lt;br&gt;
abandoned maintainer account or an expired registration, not that someone registered a fresh&lt;br&gt;
typosquat. Age alone doesn't distinguish them — &lt;code&gt;item-validator&lt;/code&gt; is, by every age-based signal,&lt;br&gt;
an old and presumably trustworthy dependency, right up until it isn't.&lt;/p&gt;
&lt;h2&gt;
  
  
  The download numbers, for what they're worth
&lt;/h2&gt;

&lt;p&gt;npm's public stats endpoint still has a record for both, from before either was pulled:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://api.npmjs.org/downloads/point/last-month/items-validator
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"downloads"&lt;/span&gt;:736,&lt;span class="s2"&gt;"start"&lt;/span&gt;:&lt;span class="s2"&gt;"2026-09-17"&lt;/span&gt;,&lt;span class="s2"&gt;"end"&lt;/span&gt;:&lt;span class="s2"&gt;"2026-10-16"&lt;/span&gt;,&lt;span class="s2"&gt;"package"&lt;/span&gt;:&lt;span class="s2"&gt;"items-validator"&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://api.npmjs.org/downloads/point/last-month/item-validator
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"downloads"&lt;/span&gt;:0,&lt;span class="s2"&gt;"start"&lt;/span&gt;:&lt;span class="s2"&gt;"2026-09-17"&lt;/span&gt;,&lt;span class="s2"&gt;"end"&lt;/span&gt;:&lt;span class="s2"&gt;"2026-10-16"&lt;/span&gt;,&lt;span class="s2"&gt;"package"&lt;/span&gt;:&lt;span class="s2"&gt;"item-validator"&lt;/span&gt;&lt;span class="o"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;736 downloads in under three weeks is not nothing for a package with two versions and no public&lt;br&gt;
repository we could find. It could mean genuine early adoption; it could mean automated download&lt;br&gt;
activity, which is a known way to make a young package look more established than it is. The&lt;br&gt;
endpoint doesn't distinguish the two, and we're not going to pick one explanation and present it&lt;br&gt;
as fact. &lt;code&gt;item-validator&lt;/code&gt;'s zero fits a package nobody had reason to install in the month before&lt;br&gt;
its old name got reused.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this fits on the pull request
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;items-validator&lt;/code&gt; is exactly what a dependency-age and adoption check is built to catch: new,&lt;br&gt;
thin version history, added to a lockfile for the first time. On a pull request that introduces&lt;br&gt;
it, that's what MagAudit marks as critical, with the file and line of the lockfile change —&lt;br&gt;
whether that stops the merge depends on whether the team has made the check required in their own&lt;br&gt;
branch protection settings. &lt;code&gt;item-validator&lt;/code&gt; is the harder case for any check built around&lt;br&gt;
novelty: a six-year-old name with a real publish history doesn't look new to an age heuristic on&lt;br&gt;
the day it's added, because in every dependency tree that already pinned it months ago, it wasn't&lt;br&gt;
added this week at all — the risk only appears retroactively, when a dormant name starts shipping&lt;br&gt;
again.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this doesn't tell you
&lt;/h2&gt;

&lt;p&gt;Neither advisory says what the payload actually did, so we haven't described one. We don't know&lt;br&gt;
whether &lt;code&gt;item-validator&lt;/code&gt;'s takeover was a compromised maintainer account, an expired email domain,&lt;br&gt;
or something else — GitHub's advisory doesn't say, and nothing in the registry metadata confirms&lt;br&gt;
it either way. We don't know if the 736 downloads of &lt;code&gt;items-validator&lt;/code&gt; represent real installs by&lt;br&gt;
real developers or inflated numbers from the campaign itself; the stats endpoint can't tell the&lt;br&gt;
two apart, and we haven't found a second source that can. And we don't know whether the two&lt;br&gt;
packages share an operator — same naming pattern, same advisory batch, same morning, but that's&lt;br&gt;
circumstantial, not confirmed.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>javascript</category>
      <category>ai</category>
    </item>
    <item>
      <title>Field report: 9917 real pull requests scanned, 7.9% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 12 Oct 2026 09:35:37 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-9917-real-pull-requests-scanned-79-flagged-46-of-our-own-mistakes-fixed-in-public-b6j</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-9917-real-pull-requests-scanned-79-flagged-46-of-our-own-mistakes-fixed-in-public-b6j</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;9917&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;7.9%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;8900&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;9293&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;11 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 9917 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-10-12 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
    <item>
      <title>How to read a package's install scripts before npm runs them</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 12 Oct 2026 08:42:53 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/how-to-read-a-packages-install-scripts-before-npm-runs-them-3e44</link>
      <guid>https://dev.to/mag_solutions_ai/how-to-read-a-packages-install-scripts-before-npm-runs-them-3e44</guid>
      <description>&lt;p&gt;You're about to add a dependency. Before &lt;code&gt;npm install&lt;/code&gt; finishes, it may already have run&lt;br&gt;
arbitrary JavaScript on your machine — not because the package is malicious, but because&lt;br&gt;
that's how npm's install lifecycle works. &lt;a href="https://docs.npmjs.com/cli/v10/using-npm/scripts" rel="noopener noreferrer"&gt;npm's own documentation&lt;/a&gt;&lt;br&gt;
lists the order: &lt;code&gt;preinstall&lt;/code&gt;, then &lt;code&gt;install&lt;/code&gt;, then &lt;code&gt;postinstall&lt;/code&gt;, all run after the package&lt;br&gt;
lands in &lt;code&gt;node_modules&lt;/code&gt; and before &lt;code&gt;npm install&lt;/code&gt; returns control to you.&lt;/p&gt;

&lt;p&gt;Most of the time that's harmless — a native module compiling itself, a CLI tool downloading&lt;br&gt;
a binary. Sometimes it's the thing a dependency-confusion or typosquat package relies on:&lt;br&gt;
the credential-stealing code we covered in &lt;a href="///blog/coinbase-wallet-npm-cluster.html"&gt;the Coinbase wallet cluster&lt;/a&gt;&lt;br&gt;
and in &lt;a href="///blog/starbucks-sdk-dependency-confusion.html"&gt;starbucks-sdk&lt;/a&gt; both ran the moment&lt;br&gt;
&lt;code&gt;npm install&lt;/code&gt; finished, via a &lt;code&gt;postinstall&lt;/code&gt; hook, with no further action from anyone.&lt;/p&gt;

&lt;p&gt;The useful part: you don't need to run the install to see the scripts. The npm registry&lt;br&gt;
stores the full &lt;code&gt;package.json&lt;/code&gt; for every published version, scripts included, and serves it&lt;br&gt;
as plain JSON.&lt;/p&gt;
&lt;h2&gt;
  
  
  The command
&lt;/h2&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/&amp;lt;package&amp;gt; | &lt;span class="se"&gt;\&lt;/span&gt;
  python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"import json,sys; d=json.load(sys.stdin); v=d['dist-tags']['latest']; &lt;/span&gt;&lt;span class="se"&gt;\&lt;/span&gt;&lt;span class="s2"&gt;
  print(d['versions'][v].get('scripts'))"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Run against five real packages, today:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/left-pad | ...
None

&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/husky | ...
None

&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/puppeteer | ...
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'postinstall'&lt;/span&gt;: &lt;span class="s1"&gt;'node install.mjs'&lt;/span&gt;, ...&lt;span class="o"&gt;}&lt;/span&gt;

&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/cypress | ...
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'postinstall'&lt;/span&gt;: &lt;span class="s1"&gt;'node dist/index.js --exec install'&lt;/span&gt;, ...&lt;span class="o"&gt;}&lt;/span&gt;

&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/core-js | ...
&lt;span class="o"&gt;{&lt;/span&gt;&lt;span class="s1"&gt;'postinstall'&lt;/span&gt;: &lt;span class="s1"&gt;'node -e "try{require(\'&lt;/span&gt;./postinstall&lt;span class="se"&gt;\'&lt;/span&gt;&lt;span class="o"&gt;)}&lt;/span&gt;catch&lt;span class="o"&gt;(&lt;/span&gt;e&lt;span class="o"&gt;){}&lt;/span&gt;&lt;span class="s2"&gt;"'}
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No install, no sandbox, no five minutes waiting for a tarball — one HTTP request per&lt;br&gt;
package.&lt;/p&gt;
&lt;h2&gt;
  
  
  Reading what the script actually does
&lt;/h2&gt;

&lt;p&gt;A &lt;code&gt;postinstall&lt;/code&gt; entry by itself tells you nothing about intent. All three packages above&lt;br&gt;
run one, and all three are doing something ordinary.&lt;/p&gt;

&lt;p&gt;Puppeteer's &lt;code&gt;install.mjs&lt;/code&gt; &lt;a href="https://unpkg.com/puppeteer@25.13.0/install.mjs" rel="noopener noreferrer"&gt;says exactly what it's for&lt;/a&gt;&lt;br&gt;
in its own header comment: "the &lt;code&gt;puppeteer&lt;/code&gt; package runs this script during the installation&lt;br&gt;
process" to fetch a supported browser build. Cypress does the same for its own binary. Both&lt;br&gt;
are well-known, well-documented reasons for a package to act during install — and both would&lt;br&gt;
trip a rule that simply flags "has a postinstall script" as a finding, which is why that rule&lt;br&gt;
on its own is not good enough.&lt;/p&gt;

&lt;p&gt;Core-js is the clearer case for actually reading the file instead of guessing from the&lt;br&gt;
command. &lt;a href="https://unpkg.com/core-js@3.50.0/postinstall.js" rel="noopener noreferrer"&gt;Its postinstall script&lt;/a&gt; checks a&lt;br&gt;
handful of environment variables — &lt;code&gt;CI&lt;/code&gt;, &lt;code&gt;ADBLOCK&lt;/code&gt;, &lt;code&gt;DISABLE_OPENCOLLECTIVE&lt;/code&gt; — and, when none&lt;br&gt;
of them are set, prints a one-time terminal banner asking for project funding, then writes a&lt;br&gt;
marker file to &lt;code&gt;os.tmpdir()&lt;/code&gt; so it won't repeat. No network call, no credential read, no file&lt;br&gt;
outside &lt;code&gt;/tmp&lt;/code&gt;. Unwelcome if you didn't expect it; not a threat.&lt;/p&gt;

&lt;p&gt;Compare that to what you'd want to see from a script in a package you've never heard of,&lt;br&gt;
published a few days ago: a request to an unfamiliar host, a read of &lt;code&gt;process.env&lt;/code&gt; passed&lt;br&gt;
anywhere outside &lt;code&gt;console.log&lt;/code&gt;, or a write to a path like &lt;code&gt;~/.aws&lt;/code&gt; or &lt;code&gt;~/.ssh&lt;/code&gt;. None of that&lt;br&gt;
appears in any of the five packages above — which is exactly the point of checking before you&lt;br&gt;
install, not after.&lt;/p&gt;
&lt;h2&gt;
  
  
  The same trick works on age
&lt;/h2&gt;

&lt;p&gt;The registry's &lt;code&gt;time&lt;/code&gt; field is in the same response, free of charge:&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;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://registry.npmjs.org/left-pad | &lt;span class="se"&gt;\&lt;/span&gt;
  python3 &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"import json,sys; d=json.load(sys.stdin); print(d['time']['created'])"&lt;/span&gt;
2014-03-14T09:09:20.762Z
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A package that's two days old with a &lt;code&gt;postinstall&lt;/code&gt; script and a handful of downloads is a&lt;br&gt;
different risk profile from one that's ten years old with a script that fetches a browser&lt;br&gt;
binary and &lt;a href="https://api.npmjs.org/downloads/point/last-week/puppeteer" rel="noopener noreferrer"&gt;11.4 million downloads in the last seven days&lt;/a&gt;&lt;br&gt;
(checked 12 October 2026). Age and adoption don't prove safety, but a script on a brand-new,&lt;br&gt;
unused name is worth five extra minutes that the same script on &lt;code&gt;puppeteer&lt;/code&gt; is not.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this doesn't tell you
&lt;/h2&gt;

&lt;p&gt;This technique reads what a script &lt;em&gt;says&lt;/em&gt; it does in source form. It does not execute the&lt;br&gt;
script in a sandbox, does not catch logic that only triggers under specific conditions, and&lt;br&gt;
does not follow code the script downloads and runs afterward — a dropper that fetches its&lt;br&gt;
real payload from a second URL at install time would still look ordinary from this check&lt;br&gt;
alone. It also doesn't tell you anything about a dependency that's one level down: these five&lt;br&gt;
packages are the ones you typed into &lt;code&gt;package.json&lt;/code&gt; yourself, not the forty-odd others that&lt;br&gt;
came along as transitive dependencies with their own scripts. And none of this is malware&lt;br&gt;
analysis — it's reading source, not behaviour; a tool built to execute packages in isolation&lt;br&gt;
and watch what they actually do, such as Socket, is answering a different and harder&lt;br&gt;
question than the one this command answers.&lt;/p&gt;

&lt;p&gt;If you want this check to run on every pull request instead of by hand, that's one of the&lt;br&gt;
things a dependency-age-and-adoption check looks at automatically and flags in the PR.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>security</category>
      <category>ai</category>
    </item>
    <item>
      <title>Field report: 6949 real pull requests scanned, 6.1% flagged, 46 of our own mistakes fixed in public</title>
      <dc:creator>Mag Solutions</dc:creator>
      <pubDate>Mon, 05 Oct 2026 09:34:11 +0000</pubDate>
      <link>https://dev.to/mag_solutions_ai/field-report-6949-real-pull-requests-scanned-61-flagged-46-of-our-own-mistakes-fixed-in-public-103g</link>
      <guid>https://dev.to/mag_solutions_ai/field-report-6949-real-pull-requests-scanned-61-flagged-46-of-our-own-mistakes-fixed-in-public-103g</guid>
      <description>&lt;p&gt;We run a security check on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the field looks like this week
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests analysed&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6949&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pull requests where the check flagged something&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6.1%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Dependencies resolved against live PyPI and npm&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6619&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distinct package names seen&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6926&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hallucinated packages confirmed&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates discarded on review&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Candidates still pending review&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sweeps, since 2026-09-04&lt;/td&gt;
&lt;td&gt;28&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A 404 means a name is not published. It does not mean anyone invented it&lt;/strong&gt; — an unpublished sibling inside a monorepo returns exactly the same 404, and no HTTP request can tell them apart. So a name is counted only once a person has confirmed it, and the discarded and pending counts are published beside the total.&lt;/p&gt;

&lt;p&gt;2 candidate(s) have not been looked at yet and are &lt;strong&gt;not&lt;/strong&gt; in the count above. Uncounted is the default: we would rather say nothing than say something we have not checked.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;None confirmed — and that is not good news for us.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A lot has been written about AI assistants inventing package names. On 6949 real pull requests we have confirmed &lt;strong&gt;none&lt;/strong&gt;, and discarded 21 candidate(s) on review — sibling packages of a monorepo, and configuration keys read as package names.&lt;/p&gt;

&lt;p&gt;We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check any of this yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The raw record: &lt;a href="https://api.magsolutionsai.com/quality" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/quality&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The live field figures: &lt;a href="https://api.magsolutionsai.com/measurement" rel="noopener noreferrer"&gt;https://api.magsolutionsai.com/measurement&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The corpus behind both, with every cited pull request: &lt;a href="https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence" rel="noopener noreferrer"&gt;https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Written up in full: &lt;a href="https://magsolutionsai.com/quality.html" rel="noopener noreferrer"&gt;https://magsolutionsai.com/quality.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Generated on 2026-10-05 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was produced automatically by our software agents (an AI model or a report template) from measurements they ran themselves, with no human editorial review before publication (EU AI Act, art. 50). Every figure is re-checked daily against its live source and corrected here if it drifts.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>opensource</category>
      <category>devops</category>
    </item>
  </channel>
</rss>
