DEV Community: The Glitch Fixer

Claude Duet: Anthropic's New Proactive AI Agent — What It Is and How to Join the Waitlist

The Glitch Fixer — Tue, 27 Oct 2026 04:41:13 +0000

Originally published at https://theglitchfixer.blogspot.com/2026/10/claude-duet-anthropics-new-proactive-ai.html

Anthropic launched a new kind of Claude on October 21, 2026: Claude Duet, a proactive AI agent now rolling out as a research preview. Instead of waiting for you to type a prompt, Duet comes to you — it sends a morning briefing drawn from your connected apps, keeps working on tasks after you close the app, and checks in only when something actually needs your attention.

It lands in the middle of an agent arms race: OpenAI Dots, Meta's Muse, and Google's agent push are all converging on the same idea. Anthropic's answer is deliberately work-first — built for professionals and small-business owners drowning in busywork, not for booking restaurant tables.

Claude Duet is a standalone proactive agent: a separate experience from regular Claude, delivered as its own apps on iPhone, Android, and the web. As Anthropic put it in its announcement: "It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf."

How Claude Duet actually works

The day starts with the briefing. Duet scans your connected email, calendar, and work apps and sends one short rundown: the items that matter, the deadlines approaching, the threads that need a reply. You can ask it to prepare draft replies, then check back every few hours for developments you asked it to flag.

The part that separates it from a chatbot is the background machine. Duet runs inside its own cloud-based computer — a virtual machine that keeps working after you close the app or shut your laptop. That is what makes the "keeps working when you step away" promise real: it is not a tab you left open, it is a separate machine doing the task.

It is designed to be interruptible on your terms. Rather than streaming everything it does, Duet judges which events deserve your attention and asks for permission before acting on your behalf. When it needs a decision — approving a payment, sending a message, signing a document — it comes back to you instead of guessing.

Under the hood, Duet runs on Claude Opus 5.5, Anthropic's flagship model, and it uses the same memories, skills, and connectors as regular Claude. It can even search your past conversations so you do not have to repeat context. Finished work lands in a Make tab for review and editing; a separate Tasks tab tracks progress, with completion left for you to mark.

How to get Claude Duet: the waitlist

Access is limited, and deliberately so. Claude Duet is a research preview — Anthropic's word for "real product, early days" — and for now it is available only through a waitlist open to Claude Pro and Max subscribers in the US. Invitations go out in waves, so joining the waitlist does not mean instant access; Anthropic has not given a timeline for how fast the waves move.

Three things to know before you join the line:

You need a paid Claude subscription. The preview is restricted to Pro and Max plans. It is included at no additional cost beyond the subscription — but read the pricing section below, because "no additional cost" comes with an asterisk.
US only, for now. The rollout starts in the United States. Team and Enterprise plans are explicitly not included yet, though Anthropic says it plans to expand to those subscriptions later.
Join through the Claude app. The waitlist lives inside the Claude experience on iOS, Android, and web — check the Claude app or Anthropic's announcements for the current signup entry point, since preview access points can move during a rollout.

Is Claude Duet free?

The feature itself costs nothing extra: it ships inside your existing Pro or Max subscription. The catch is usage. Anthropic says Duet counts toward your plan's existing usage limits — including work it performs in the background — and warns that this could make you reach those limits "sooner than with chat alone." A busy agent silently burning quota overnight is the most practical cost of the preview, so keep an eye on your usage dashboard if you get in.

What Claude Duet connects to

Anthropic says Duet works with more than 6,000 connectors and plugins through the Claude Marketplace — the same connector ecosystem as regular Claude — including:

Email and messaging: Gmail and Slack for the morning brief, thread monitoring, and drafted replies.
Money: Stripe (spot unpaid invoices and draft payment reminders), QuickBooks, and Gusto (prepare payroll information for review).
Sales: HubSpot for follow-ups based on a deal's history.
Professional network: LinkedIn.
Commerce: Shopify — Duet can even complete purchases through Shop Pay, with your approval.

Beyond connections, Duet can build things: financial and sales forecasts, dashboards, prototypes, and web apps. Small-business examples Anthropic shared include accounting, customer support, marketing, and invoices across multiple locations.

The safety rails: what it asks permission for

Duet is designed to ask before it sends a message, makes a payment, signs a document, or deletes anything. The approval step is the product's defining trait — the agent proposes, you dispose.

There is also a new layer of automated screening. Anthropic released a classifier that screens each message to stop sensitive information from leaking out: by default, Duet turns down requests involving credit card numbers and government IDs. And some actions are off-limits entirely — Duet will not trade securities or transfer funds between accounts, even if you approve them.

On the access side, Anthropic says Duet does not directly touch your passwords or sign-in tokens; it works through the connected apps' official integrations rather than logging in as you. The honest caveat, from Anthropic itself: the agent "might misread a thread or miscalibrate what's urgent." An approval gate is only as good as the judgment behind it, which is why the sensitive actions always route through you.

Claude Duet vs OpenAI Dots vs Meta Muse

OpenAI Dots — announced at DevDay on September 29, always-on agents built on GPT-6 Astra with 4,000+ app plugins and desktop-based creation. Dots leans toward everyday assistance and personal errands.
Meta Muse — Meta's frontier personal agent, pitched around daily life and personal tasks.
Claude Duet — the work-first entry: morning briefings, background execution on its own virtual machine, 6,000+ connectors skewed toward business apps, and an explicit "asks before it acts" approval model.

The most telling difference is one Anthropic volunteers: Duet is weaker at personal errands like booking tickets than its rivals. Dots and Muse want your daily life; Duet wants your workday.

The catches you should know

It is a research preview. Expect bugs, missing features, and behavior changes without warning. You cannot watch its browser use as it happens yet, so background work is a black box until it reports back.
Background work burns quota. Everything Duet does — including the tasks it runs while you sleep — counts against your Pro or Max usage limits. Anthropic's own warning: you may hit your limits "sooner than with chat alone."
Access is narrow. US only, Pro/Max only, waitlist only, invitations in waves. Team and Enterprise customers wait. No public timeline for expansion.
It can misjudge urgency. Anthropic cautions that Duet "might misread a thread or miscalibrate what's urgent" — the morning brief is a starting point, not a verdict.
Personal errands are the weak spot. If your dream agent books flights and grabs concert tickets, Duet is not that agent yet. Its rivals cover that ground better today.

Should you join the waitlist?

If you are already paying for Claude Pro or Max and your work lives in Gmail, Slack, Stripe, or HubSpot, yes — join the waitlist. The morning brief plus background execution is a genuinely new capability, and as a research preview it costs nothing beyond the subscription you already have. Just watch your usage limits once you are in.

If you are not a subscriber, or you want an agent for personal errands rather than work, there is no reason to rush. Duet's work-first design is its identity; the broader agent race will give you more options soon.

Frequently asked questions

What is Claude Duet? Anthropic's proactive AI agent, announced October 21, 2026 as a research preview. It sends you a morning briefing from your connected apps, works on tasks in the background on its own cloud computer, and checks in only when something needs your attention — asking permission before it sends messages, makes payments, signs documents, or deletes anything.

How do I get access to Claude Duet? Join the waitlist inside the Claude app on iOS, Android, or web. You need an active Claude Pro or Max subscription and a US account; Anthropic sends invitations in waves, so there may be a wait. Team and Enterprise plans are not included yet.

Is Claude Duet free? The preview is included at no extra cost with a Claude Pro or Max subscription. But all of Duet's work — including background tasks — counts toward your plan's usage limits, and Anthropic warns you may reach those limits sooner than with chat alone.

What apps can Claude Duet connect to? More than 6,000 connectors via the Claude Marketplace, including Gmail, Slack, LinkedIn, HubSpot, Stripe, QuickBooks, Gusto, and Shopify. It can also build forecasts, dashboards, prototypes, and web apps.

Can Claude Duet make purchases or payments? It can complete purchases through Stripe's Link wallet and Shopify's Shop Pay, but only with your explicit approval each time. It refuses to share credit card numbers or government IDs, and it will not trade securities or transfer funds between accounts under any circumstances.

How is Claude Duet different from OpenAI Dots? Dots leans toward everyday personal assistance, while Duet is deliberately work-first: morning briefings, background execution for professionals and small businesses, and an explicit ask-before-acting approval model. Anthropic admits Duet is weaker at personal errands like booking tickets.

KSMBD Vulnerability: Your NAS Runs Samba — But Is It the Kernel One?

The Glitch Fixer — Tue, 27 Oct 2026 04:38:14 +0000

Originally published at https://thehomelabfixer.blogspot.com/2026/10/ksmbd-vulnerability-your-nas-runs-samba.html

Two remote code execution flaws rated 9.0 out of 10 were disclosed on October 22, 2026, in Linux's in-kernel SMB server — and the first thing you need to know is that they probably do not affect you at all. Most homelabs share files with normal Samba, which runs in userspace. These flaws are in KSMBD, the kernel's own built-in SMB server, which is a different thing entirely. This post gives you a 30-second check to find out which one your NAS runs, what to switch off if it is the kernel one, and what to watch for the fix.

Honesty first: as of today (Oct 26), these flaws have no CVE numbers and no confirmed fix in a stable or distribution kernel. They are tracked as ZDI-CAN-33774 and ZDI-CAN-33549 by Trend Micro's Zero Day Initiative, which rated both 9.0. The kernel project only assigns CVE numbers once a stable-tree fix exists, so until then, ZDI-CAN IDs are the correct references. Unverified secondary reports mention a fix in a branch aimed at Linux 7.5, but it is not confirmed and it is unclear which flaw it covers — so this post does not count on it.

The 30-second check: are you affected?

Run one of these on your NAS or Linux server:

systemctl status ksmbd
lsmod | grep ksmbd

If neither shows KSMBD running, stop here — you are not exposed to these two flaws. If you run normal Samba (the smb/smbd service most NAS systems and distros use), you are also not affected. Only a system with the KSMBD kernel module active is in scope.

Two more facts that narrow it further. First, KSMBD is not enabled by default on many distributions — Red Hat Enterprise Linux, for example, does not include it by default. Second, the two flaws have very different exposure:

ZDI-CAN-33774 — the dangerous one. A race in how KSMBD assigns its handler table: a burst of short SMB1 packets followed by an SMB2 packet in the same TCP write makes workers handle SMB1 data with an SMB2 handler, reading past the end of the shorter buffer. A path to writing past the buffer also exists once an attacker has access to a share. It needs no authentication and exposes any server running KSMBD on port 445.
ZDI-CAN-33549 — the narrow one. An integer underflow in the SMB Direct path: a client supplies RDMA buffer descriptors where the first is longer than the declared payload, the server's size counter wraps, and a later memory copy overruns an allocated buffer. It needs three things at once: an RDMA-capable interface (hardware, or the rxe software driver), a writable KSMBD share, and a client negotiating SMB Direct. Most homelabs meet none of those, let alone all three.

So the real question is not "is Linux vulnerable" — it is "is KSMBD running on your box, and is port 445 reachable?" Answer those two and you know your risk.

If KSMBD is running: lock it down now

These are interim mitigations until a real fix lands in your distro's kernel — not vendor-certified guidance, because no vendor fix exists yet.

Stop the KSMBD service and unload the module if you do not need kernel SMB. (Commands vary by distro — check your distro's documentation. If the module is built into your kernel rather than loadable, the firewall step below is your control.)
Block port 445 inbound at your firewall for anything that does not strictly need SMB from outside your LAN. For most homelabs, SMB should never face the internet at all — this step alone removes the unauthenticated flaw's reach.
Disable SMB Direct if you use it, unless you have a specific RDMA need — that closes the second flaw's path.
Switch file sharing to userspace Samba if your setup allows it. Normal Samba is a different codebase and is not affected by these two flaws.
Check your kernel and distro version and note it down, so you recognise the fixed version when it arrives.

Do not confuse these with the other flaws disclosed in the same batch — a local privilege escalation in the networking stack, one in a virtual GPU driver, an information disclosure in an NVMe driver, and a buffer overflow in the ConnMan daemon. Those are separate issues with different fixes; this post is only about the two KSMBD flaws.

Fix-watch: what to look for

Watch three things, in this order:

A fix landing in the stable kernel tree, then in your distribution's kernel (fixes typically take weeks to travel from mainline to stable branches and distro kernels).
CVE numbers being assigned to ZDI-CAN-33774 and ZDI-CAN-33549 — that assignment itself is the signal a stable fix exists.
Your distro's security advisory naming the fixed kernel version — that is the version to upgrade to.

The original post will be updated when the fixed version is confirmed. Until then, the check commands above plus a blocked port 445 are your protection.

Bottom line

Your NAS probably runs Samba, not KSMBD — and if it does, these two 9.0s are not your problem. Thirty seconds with one command tells you for sure. If it turns out you ARE running the kernel server, stop it or firewall it today; the unauthenticated flaw needs nothing more than a reachable port 445.

FAQs

Does the KSMBD vulnerability affect my Samba NAS? Only if your NAS runs KSMBD, the kernel's built-in SMB server — not normal userspace Samba. Run systemctl status ksmbd or lsmod | grep ksmbd: if neither shows KSMBD active, you are not affected.

Is there a CVE number or patch for the KSMBD flaws? Not as of Oct 26, 2026. The flaws are tracked as ZDI-CAN-33774 and ZDI-CAN-33549. The kernel project assigns CVE numbers only after a stable-tree fix exists, and no confirmed stable or distribution fix has landed yet.

Which of the two KSMBD flaws is more dangerous? ZDI-CAN-33774 — it needs no authentication and exposes any KSMBD install on port 445. ZDI-CAN-33549 needs an RDMA interface, a writable share, and SMB Direct negotiation all at once, so its exposure is far narrower.

How do I protect my server until a fix lands? Stop the KSMBD service and unload the module if you don't need kernel SMB, block inbound port 445 at your firewall, disable SMB Direct if you use it, or switch file sharing to userspace Samba.

I run Red Hat Enterprise Linux. Am I affected? RHEL does not include KSMBD by default, so a stock RHEL system is not running it. Run the check commands above to confirm nothing has enabled it on your system.

Xbox Game Pass New Games: 50 Titles Arrive October 26–30, Only 3 Are on Game Pass

The Glitch Fixer — Tue, 27 Oct 2026 04:33:56 +0000

Originally published at https://theglitchfixer.blogspot.com/2026/10/xbox-game-pass-new-games-50-titles.html

Xbox Wire, Xbox's official news site, published its "Next Week on Xbox" roundup on October 23, 2026. It lists 50 new games for October 26 to 30, 2026. We counted the list ourselves: 3 + 9 + 9 + 20 + 9 = 50.

Two facts matter most. Only three of the 50 carry the Game Pass tag in Xbox Wire's list. And 20 of the 50 release on a single day, October 29. Everything below comes from Xbox Wire unless we clearly say otherwise. Where we give an opinion, we label it as our take.

The only 3 Game Pass titles this week

These are the three titles Xbox Wire lists with Game Pass:

Scramble Knights Online (Game Preview) — October 26. Game Preview is Xbox's early-access program, which means the game is still in development.
The Séance of Blake Manor — October 27.
Terrible Lizards — October 29.

That is the complete Game Pass list in this roundup. We do not call them "day one" titles, because Xbox Wire's list does not use that label — it tags them with Game Pass, and that is exactly what we report.

October 29: 20 games in one day

October 29 carries 20 of the 50 games, including:

Mouthwashing — October 29, $12.99. Google Trends shows "mouthwashing xbox" as a rising search (+400%) over the past 7 days.
The Wolf Among Us Remastered — October 29. Note: some sites date this October 26. Xbox Wire says October 29, and we follow Xbox Wire.
Croc 2 Kingdom of the Gobbos — October 29.
Hunting Simulator 3 — October 29, $49.99.
Terrible Lizards, listed above as a Game Pass title, also releases October 29.

Prices and details above are only the ones Xbox Wire's list states. Where a price or description is not given here, it is because the source list does not state it — we did not fill gaps with guesses.

The full list: all 50 Xbox new games, day by day

October 26 (3): Scramble Knights Online (Game Preview) — Game Pass; Ayasa: Shadows of Silence; PAO PAO (Game Preview).

October 27 (9): The Séance of Blake Manor — Game Pass; Bao Bao's Cozy Cleaning Services; Bloom Busters; Cozy Radiation Cleaner; Dog Man: Bark in Action; Farming Simulator 25: Beans & Alpacas Edition; Remothered: Red Nun's Legacy; Sneaky Purr; Treasure Tiger.

October 28 (9): Fall Adventure; Fallen City Brawl; Gennady; House Builder 2; Karate Survivor; The Last Case; Platypus Steps; Primal Planet; SULFUR.

October 29 (20): Mouthwashing; The Wolf Among Us Remastered; Beyond The Evil; Candy Shop Simulator; The Cat in the Hat: Rainy Day Mayhem; Choice of Life: Floors; Choice of Life: Samosbor; Croc 2 Kingdom of the Gobbos; Curse of Resthaven; HELLBOUND INC; Hunting Simulator 3; HYPERVIOLENT; Joe & Mac Retro Collection; Love: Fall in Death; My Life: Zoo Vet; Planted!; The Transference; Train Yard Builder; The Zibbo Show; Terrible Lizards — Game Pass.

October 30 (9): Queen's Domain; The Abjuration Stand; Cat Pipes: More Meow; Corner Shop: NightShift; Ex-Master, Zombie; Love Love Mystery Club; The Night of the Exorcism; Pinballoid Crush; Space Rock Breaker.

Our take: what to play first

Everything in this section is our opinion, not news.

If you already have Game Pass, start with the three Game Pass titles — they are the week's no-extra-cost picks. Among them, our pick for story is The Séance of Blake Manor, and for action, Terrible Lizards.

If you are buying one game, our pick is Mouthwashing. At $12.99, it is the cheapest of the headliners with a confirmed price, and it is the one searches are rising for.

One practical warning: the roundup spans Xbox Series X|S, Xbox One, and PC, and not every game is on every platform. Check the game's store page for your console before you buy.

What this list is not

It is not October 26–31. Some sites stretch the window by a day. Xbox Wire's list ends October 30, and every date above follows Xbox Wire.

It is not 50 Game Pass games. Only the three named above carry the Game Pass tag in this roundup.

It is not a review. We have not played these games. This is verified release news, plus our take, clearly marked.

It is not complete with genres and prices for every game. Xbox Wire's list gives titles and dates. We report only what it states, and we say so where information is missing, rather than inventing it.

Source: Xbox Wire, "Next Week on XBOX: New Games for October 26 to 30" (October 23, 2026) — https://news.xbox.com/en-us/2026/10/23/next-week-on-xbox-new-games-for-october-26-to-30/

FAQs

How many new Xbox games release October 26–30, 2026? 50, counted from Xbox Wire's own list: 3 on October 26, 9 on October 27, 9 on October 28, 20 on October 29, and 9 on October 30.

Which new Xbox games are on Game Pass this week? Three titles Xbox Wire lists with Game Pass: Scramble Knights Online (Game Preview) on October 26, The Séance of Blake Manor on October 27, and Terrible Lizards on October 29.

Are all 50 new Xbox games on Game Pass? No. Only the three above carry the Game Pass tag in Xbox Wire's list. The other 47 are regular releases.

When does Mouthwashing release on Xbox? October 29, 2026, priced at $12.99, per Xbox Wire.

When does The Wolf Among Us Remastered release on Xbox? October 29, 2026, per Xbox Wire. Some sites say October 26, but Xbox Wire's date is October 29.

Node.js Update: Patch the Node2Shell RCE Tonight — Self-Hosted Fix Guide (CVE-2026-111497)

The Glitch Fixer — Sat, 24 Oct 2026 03:13:12 +0000

The only sane advice this week: patch tonight. On Wednesday, October 21, 2026, Node.js shipped security releases for the 26.x, 24.x and 22.x lines. By Thursday, a working proof of concept was public — researchers had demonstrated remote code execution on the official Node.js 24.22.0 Docker image using crafted Brotli data. Today, The Hacker News carried the story wider. The vulnerability has a nickname now: Node2Shell.

If you self-host, this one matters more than most Node advisories. Node.js isn't just a runtime on your server — it is inside your containers. Your dashboards, automation and media tooling ship a Node runtime that a host package audit won't find. This guide is a runbook: find every Node in your stack, then fix each one — including the three traps that catch self-hosters.

Honest-testing note: We did not reproduce any exploit — no exploit or PoC instructions here, only defensive checks. Every fact below is quoted or derived from the official Node.js advisory, Imperva's research, Singapore's CSA advisory, or The Hacker News' coverage.

What Node2Shell actually is

Node2Shell is CVE-2026-111497, an out-of-bounds write in Node.js's bundled Brotli decoder. The official advisory: "A flaw in Node.js can cause Node.js Brotli decoder out-of-bounds write permits remote code execution." It affects the 26.x, 24.x and 22.x lines; the fixed builds are 26.12.1, 24.22.1 and 22.23.4. Reported by OpenAI Codex Security — Khai Tran and Sylvie Mayer; fixed by mcollina and aduh95.

Two things to get right on severity. Node.js's own words: "The highest severity issue fixed in this release is HIGH." Separately, Singapore's CSA scored the same CVE at CVSS v3.1 9.8 — the CSA's score, not the Node.js project's rating. Per CSA: "Successful exploitation of this out-of-bounds write vulnerability could allow an attacker to execute arbitrary code remotely or disclose sensitive information from the memory of affected systems. Exploitation requires an affected application to decompress attacker-controlled Brotli-compressed data."

Imperva's research identified three paths: (1) incoming request bodies with Brotli encoding — Express 5 (versions 5.0.0–5.3.0) and body-parser 2.x (2.0.0–2.3.0) decompress these; (2) outbound fetch() calls decoding Brotli responses from servers you don't control; (3) direct use of node:zlib. Imperva demonstrated RCE against the official Node.js 24.22.0 Linux arm64 image and confirmed the fixed 24.22.1 rejected the same input. No in-the-wild exploitation was reported as of October 22 — a window-of-opportunity story, which is exactly when patching pays off.

This release also fixed ten vulnerabilities total (four high, three medium, three low) — TLS hostname-check bypasses, a dns.resolve NUL-truncation bug, and bundled updates (c-ares, ca-cert with the eMudhra Intermediate CA compromise, OpenSSL, SQLite, Undici). More reasons to update Node.js everywhere it hides.

Step 0 — Are you exposed? Audit every Node in your stack

Start with the Node you know about, on the host:

node -v
# If it prints v22.23.4, v24.22.1 or v26.12.1 (or newer), the host runtime is fixed.

Then the Node you forgot about — inside your containers. A package inventory misses a runtime bundled in an image, so query the image itself:

docker run --rm my-app-image node --version
# Repeat for every Node-based app image in your compose stacks

Critical detail: check the running container, not just the pulled image — an old running container survives an image update:

docker exec my-app-container node --version
# or, for compose: docker compose exec app node --version

Make a list: every container older than 22.23.4 / 24.22.1 / 26.12.1 on its line, plus anything on an end-of-life line. That's your patch queue for tonight.

Step 1 — Re-pull and verify (the image-publisher lag trap)

The lag trap: Your app's Docker image is not fixed until its maintainer rebuilds it on node ≥22.23.4 / ≥24.22.1 / ≥26.12.1. Updating your image tag or restarting the container changes nothing if the publisher hasn't rebuilt yet. Never assume a re-pull fixed it — verify.

Before touching anything, snapshot first — the same snapshot-first discipline as our SafePay ransomware NAS survival guide. Then, for each Node-based app in your patch queue:

cd /path/to/your/compose/stack
docker compose pull          # get the publisher's latest build
docker compose up -d         # recreate containers on the new image
docker compose exec app node --version   # verify the RUNNING container is fixed

Two failure modes. First, a cached digest: if the version still shows old after the pull, confirm the digest actually changed (docker images --digests) — pinned digests and cached layers redeploy the old image silently. Second, the publisher may not have rebuilt yet — check the project's release notes. Same patch-tonight discipline as our GitLab CVSS-10 RCE guide: the vendor's fix and your deployment of it are two different events.

Step 2 — Fix the pin trap

If you build your own images, check your base-image pins. Aikido's warning, in spirit: a Dockerfile pinned to node:24.22.0 stays vulnerable no matter how often you rebuild it — rebuilding on the old pin changes nothing. So:

# In your Dockerfile, change:
#   FROM node:24.22.0-alpine
# to the fixed patch on your line:
#   FROM node:24.22.1-alpine
# Then rebuild:
docker build --no-cache -t my-app:latest .
docker run --rm my-app:latest node --version

Move to 24.22.1, 22.23.4 or 26.12.1 (or later) on whichever line you run. Same careful sequencing as our n8n 3.0 breaking-changes walkthrough — line jumps need release-note reading; patch bumps within a line are tonight's safe, fast move.

Step 3 — The EOL trap

The official advisory is blunt about end-of-life lines: "It's important to note that End-of-Life versions are always affected when a security release occurs." Per third-party guides (Aikido and others), lines like 20.x and earlier, plus odd lines 23.x/25.x, receive no fix — that detail comes from third-party analysis, not the official page.

The homelab killer: check your image's base. If any image you run is still built on an EOL Node line, no re-pull will ever fix it. Two moves: migrate the app to an image built on a supported line (22.x or 24.x), or — if the app can't move tonight — pull it off the public internet and sandbox it behind your VPN. Verify each project's Dockerfile or release notes yourself rather than trusting memory.

Can't patch tonight? A bridge-only stopgap

The official advisory offers no workaround in Node.js itself — patching is the fix. For the container you genuinely can't update tonight, reduce inbound exposure at the reverse proxy by rejecting Brotli request bodies. A bridge only: it does nothing about Brotli responses your app fetches, so don't treat it as protection.

For Nginx Proxy Manager, add to the proxy host's Custom Nginx Configuration:

# Node2Shell stopgap — TEMPORARY, remove after patching
if ($http_content_encoding ~* "br") {
  return 415;
}

For Caddy (in the site's handle block):

# Node2Shell stopgap — TEMPORARY, remove after patching
@brotli header Content-Encoding br
respond @brotli 415

Legitimate browsers sending Brotli request bodies will get a 415 — acceptable for one night on most homelab services, since Brotli request bodies are rare. Schedule the real node js security update for the morning and remove the rule once you're patched.

A note on Bun

If any self-hosted app runs on Bun rather than Node.js: Bun before 1.4.4 is affected, and Bun 1.4.4 includes the fix (per Aikido's FAQ). bun --version on the host or in the container tells you where you stand — update it the same way you'd update Node.js.

Verify the fix

Close the loop the same way you opened it — confirm the running state, not your intent:

# Host runtime
node -v
# Every patched container, checked live:
docker compose exec app node --version
docker exec my-other-app node --version

Expected: 22.23.4, 24.22.1 or 26.12.1 (or newer) on each app's line. Anything still old goes back on the queue — or gets the stopgap and an off-internet sandbox until it can be patched.

One last operator caveat: if you run Watchtower or similar auto-updaters, exclude or pin images prone to major version jumps — the same discipline every nodejs docker update demands — so tonight's fix doesn't arrive as tomorrow's breaking change. Patch the runtime, verify the running container, then sleep.

Companion read: our Plex2Shell compromise check and patch guide covers the same patch-tonight workflow for Plex servers.

Node2Shell FAQ — quick answers

Which Node.js versions fix Node2Shell (CVE-2026-111497)?

The October 21, 2026 releases: 22.23.4, 24.22.1 and 26.12.1 for the 22.x, 24.x and 26.x lines. End-of-life lines (per third-party guides, 20.x and earlier) get no fix — the official advisory states end-of-life versions are always affected.

Is Node2Shell being exploited in the wild?

No in-the-wild exploitation was reported as of October 22, per The Hacker News — but a working proof of concept was publicly demonstrated on October 22. Patch now, while the window is open.

How do I check if my Docker containers run a vulnerable Node version?

Query the image directly — docker run --rm <image> node --version — since a host package audit misses runtimes bundled inside images. Then verify the running container with docker exec <container> node --version, because an old running container survives an image update.

My app image still reports the old Node version after updating — what now?

Three checks: (1) confirm the image digest actually changed after the pull — a cached or pinned digest redeploys the old image; (2) check the project's release notes — the publisher may not have rebuilt on the fixed base yet; (3) check your Dockerfile pins — a Dockerfile pinned to node:24.22.0 stays vulnerable no matter how often you rebuild it.

Does blocking Brotli at my reverse proxy fully protect me?

No. Rejecting Content-Encoding: br request bodies at the proxy only covers one attack path. It does nothing about Brotli responses your app fetches from servers you don't control, or direct node:zlib use. A temporary bridge for the night — the real fix is the Node.js update.

Is Bun affected by Node2Shell?

Bun before 1.4.4 is affected; Bun 1.4.4 includes the fix (per Aikido's FAQ). Run bun --version on the host or in the container and update the same night.

Originally published on The Homelab Fixer.

GitLab Update: Patch the CVSS-10 RCE Tonight — Self-Hosted Fix Guide (CVE-2026-110264)

The Glitch Fixer — Fri, 23 Oct 2026 05:53:02 +0000

Originally published on The Homelab Fixer.

On October 21, 2026, GitLab shipped a critical patch release: 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10, and 18.11.13 for Community and Enterprise Edition. The headline fix is CVE-2026-110264 — an integer overflow in the regex compiler that, in GitLab's own words, "could have allowed an unauthenticated user to execute arbitrary code on the GitLab server." Severity: Critical, CVSS 10.0. Every self-hosted GitLab from version 7.0 onward is exposed unless it's on a fixed build.

This is the operator's version of the story: which build your line needs, how to back up first, the Docker and Omnibus upgrade path, the migration-downtime trap on 19.4.x, and what to do if you can't patch tonight. Same patch-tonight discipline as our php2shell emergency RCE playbook — different stack, same urgency.

⚠️ Honest testing note. We did not reproduce an exploit in a lab — deliberately. Every fact below is quoted or derived from GitLab's official patch release page and standard documented upgrade procedures. Commands are the documented forms; adapt hostnames, container names, and paths to your setup.

What CVE-2026-110264 Actually Is

Strip the CVE-speak and the bug is this: GitLab compiles regular expressions with the Onigmo engine, and a specially crafted regex could trigger an integer overflow during compilation — handing an attacker who never logged in a path to remote code execution on your server. The CVSS vector says it all: network attack, low complexity, no privileges, no user interaction, full impact on confidentiality, integrity, and availability. Reported by joaxcar through HackerOne's bug bounty program.

Two things make this worse than a normal critical patch: the version range is enormous — every self-hosted GitLab from version 7.0 onward that hasn't patched is exposed, including the classic homelab "it just works" box; and the patch carries database migrations. GitLab warns: "This patch includes database migrations that may impact your upgrade process." On single-node installs — nearly every homelab — "this patch will cause downtime during the upgrade as migrations must complete before GitLab can start." Patching is not a 30-second restart this time. Plan for it.

Step 0: Are You Vulnerable? (60-Second Check)

Get your exact version first — two reliable ways:

# Omnibus install
sudo gitlab-rake gitlab:env:info | grep -i "gitlab version"

# Docker install
docker exec gitlab gitlab-rake gitlab:env:info | grep -i "gitlab version"

Or open your instance's Help page (https://your-gitlab/help) — the version is printed at the top. Then match it against your line's fixed build:

Your line	Patch to (fixed)
19.5.x	19.5.1
19.4.x	19.4.3
19.3.x	19.3.6
19.2.x	19.2.9
19.1.x	19.1.9
19.0.x	19.0.10
18.11.x or older	18.11.13 (then plan the multi-stop upgrade below)

Anything below its line's fixed build is exposed — including instances upgraded earlier this month. Search interest is currently spiking for specific versions like 19.4.2: if that is you, you are exactly one patch behind the fix.

✅ The rule: stay on your current minor line and take its patch build. Do not jump minor versions in the same move — GitLab's upgrade path rules still apply, and skipping lines is how upgrades break.

Step 1: Back Up Before You Touch Anything

Migrations plus a security patch is the highest-risk upgrade combination. Back up before upgrading, with the snapshot-first discipline from our SafePay NAS survival guide:

# Omnibus — the documented backup task (standard procedure)
sudo gitlab-backup-cli backup create

# Docker — run the backup inside the container, then copy it off the host
docker exec -t gitlab gitlab-backup-cli backup create
docker cp gitlab:/var/opt/gitlab/backups/ /safe/place/off/the/host/

Also snapshot the VM or the Docker volumes if you can. Verify the backup file actually exists and has a sane size before proceeding — an untested backup is a rumor, not a backup.

Step 2: Upgrade — Docker

Most homelabs run GitLab in Docker. The procedure is pull, stop, remove, recreate — with one critical detail:

# Pull the EXACT fixed tag for your line (example: 19.5.x line)
docker pull gitlab/gitlab-ce:19.5.1-ce.0

# Stop and remove the old container (data lives in volumes, not the container)
docker stop gitlab
docker rm gitlab

# Recreate with your ORIGINAL run flags — same volumes, same ports, same hostname
docker run --detach \
  --hostname gitlab.example.com \
  --publish 443:443 --publish 80:80 \
  --name gitlab \
  --restart always \
  --volume /srv/gitlab/config:/etc/gitlab \
  --volume /srv/gitlab/logs:/var/log/gitlab \
  --volume /srv/gitlab/data:/var/opt/gitlab \
  gitlab/gitlab-ce:19.5.1-ce.0

Three homelab-specific warnings: pin the tag, not :latest — :latest can pull a new minor line and break your upgrade path. Watchtower users: exclude GitLab or pin it — an unattended major-version jump on a database-backed app is how instances get stranded (same upgrade discipline as our n8n 3.0 breaking-changes guide). And expect a long first start — the container runs migrations before the web UI responds, so don't restart it mid-migration.

Step 3: Upgrade — Omnibus (apt/yum)

Bare-metal and VM installs use the package manager. Pin the version so you land on the fixed build, not whatever is newest:

# Debian/Ubuntu — install the exact fixed build for your line
sudo apt update
sudo apt install gitlab-ce=19.5.1-ce.0

# Then reconfigure (standard procedure — applies the new version)
sudo gitlab-ctl reconfigure

On RHEL-family systems the equivalent is sudo yum install gitlab-ce-19.5.1-ce.0 (or dnf), then the same reconfigure. The reconfigure runs the migrations — on a single-node box, GitLab stays down until they finish. That is expected and documented, not a failure.

Step 4: Upgrade — Helm (Kubernetes)

Helm-chart installs: upgrade to the chart version that maps to your line's fixed app build (helm upgrade with the target chart version from GitLab's Helm chart release notes). Same migration caveat applies — watch the migration jobs complete before declaring victory. One paragraph is all this path needs; the chart's own upgrade notes are the authority here.

Step 5: Verify — Don't Trust the Banner

After the upgrade, verify properly — GitLab's in-app "update available" banner is a lagging indicator:

Re-run the version check from Step 0 — you should see your line's fixed build.
Open Admin > Monitoring > Background Migrations and confirm every migration shows finished. On 19.4.x there are both regular and post-deploy migrations — the instance isn't fully patched until both are done.
Check the logs for Onigmo-related crashes around the upgrade window. GitLab's published detections — "Onigmo native worker crash" and "Org mode render error burst" — double as a post-upgrade sanity signal.

Older than 18.11? You can't jump straight to a fixed build — GitLab requires stepping through intermediate versions, so this becomes a planned maintenance window, not a tonight task. Start it now, and apply the deferred-upgrade hardening above in the meantime.

Can't Patch Tonight? Reduce Exposure Now

If the upgrade has to wait for a maintenance window, shrink the attack surface: take GitLab off the public internet (Tailscale/WireGuard or IP allowlist — an unauthenticated RCE can't reach what it can't route to); restrict public projects and open sign-ups; deploy GitLab's detection rules ("Onigmo native worker crash", "Org mode render error burst"); and if you suspect compromise, rotate secrets — CI/CD variables, runner tokens, deploy keys — then patch, then rotate again.

None of this replaces the patch. GitLab's recommendation is unambiguous: upgrade affected installations "as soon as possible." The hardening buys you a safer window, not a pass.

FAQ

What is CVE-2026-110264?

An integer overflow in GitLab's Onigmo regular expression compiler that could let an unauthenticated attacker execute arbitrary code on the GitLab server. Rated Critical, CVSS 10.0, reported by joaxcar via HackerOne, and fixed in GitLab 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10, and 18.11.13 (released October 21, 2026).

Am I affected if I run GitLab in Docker?

Yes — when no specific deployment type is mentioned in GitLab's advisory, all types are affected, including Docker, Omnibus, source, and Helm installs. Check your version with docker exec gitlab gitlab-rake gitlab:env:info and compare against your line's fixed build.

Will the upgrade cause downtime?

On single-node installs, yes — GitLab states the patch "will cause downtime during the upgrade as migrations must complete before GitLab can start," particularly on the 19.4.x and 19.3.x lines which include regular and post-deploy migrations. Multi-node installs can use zero-downtime procedures. Schedule accordingly.

I upgraded GitLab earlier this month. Am I safe?

Only if your build is at or above your line's fixed version (19.5.1 / 19.4.3 / 19.3.6 / 19.2.9 / 19.1.9 / 19.0.10 / 18.11.13). Anything below those lines — including builds from earlier in October — is still exposed to CVE-2026-110264. Re-check your exact version; don't assume.

Should I use :latest for the Docker image?

No. Pin the exact fixed tag for your line (e.g. gitlab/gitlab-ce:19.5.1-ce.0) — :latest can pull a new minor version and break GitLab's upgrade path. If you use Watchtower, exclude GitLab or pin it.

How do I know the patch actually applied?

Re-run the version check, then go to Admin > Monitoring > Background Migrations and confirm all migrations show finished — don't rely on the in-app update banner. On 19.4.x, verify both regular and post-deploy migrations completed.

Character.AI Shutdown: Save Your Chats + 5 Alternatives

The Glitch Fixer — Thu, 15 Oct 2026 11:10:56 +0000

Character.AI Shutdown: Save Your Chats + 5 Alternatives

On October 12, 2026, Character.AI — the chatbot platform where millions of people build and chat with AI personas — filed for Chapter 11 bankruptcy. The company says the app keeps running for now, but warns plainly: if no buyer steps in by early December 2026, the platform will be wound down.

If you've spent months building characters and long-running conversations, this is the moment to act. Below: what the filing means for your chats, characters, and subscription, the 5 things to do right now, and the 5 best verified alternatives if the service disappears.

Is Character.AI Shutting Down?

Not yet — but the clock is ticking. The confirmed timeline, from Character.AI's own blog post and court filings:

October 12, 2026: Character Technologies, Inc. filed a voluntary Chapter 11 petition in the U.S. Bankruptcy Court for the District of Delaware, listing liabilities between $50 million and $100 million.
Right now: The service "remains up and running" and paid subscriptions stay active with billing unchanged.
Late November – early December 2026: The court-supervised sale must complete. Disney is lined up as the first ("stalking horse") bidder at $20 million for certain assets plus a non-exclusive license to the technology — not a bid to run the platform as-is.
If no buyer keeps it running: Character.AI says "the platform will be wound down."

Why did it happen? CEO Colin Adams cited "significant legal, regulatory, and operational headwinds" — lawsuits over chatbots' effects on teenagers, state investigations, and the brutal economics of running AI models for millions of daily chatters (creditors include Google, DigitalOcean, and Fireworks.ai).

What Happens to Your Chats, Characters & Subscription

Your chats and characters: Nothing changes for now. But if the platform winds down, conversation history and character definitions stored only on Character.AI's servers could disappear. No public guarantee of a migration path exists.

Your data: The company says your information stays protected under its current privacy policy during the sale. If user data becomes part of a sale, U.S. bankruptcy rules can require court approval and a privacy ombudsman review before personal data changes hands.

Your subscription: Paid plans remain active, billing unchanged. The company promises advance notice of changes. If you pay annually, refunds after a wind-down are not guaranteed — weigh your remaining prepaid time.

Do This NOW: 5 Steps Before Anything Happens

1. Export your data today (don't wait for December)

Character.AI documents a self-service data export — run it now while the service is fully working:

On the web:

Open your Profile.
Click Profile Settings.
Go to Account → Manage Account & Data.
Select Export data and follow the prompts.

In the mobile app: Profile → settings icon → Export my data.

The export delivers an archive (users report a ZIP of JSON files). Open it and confirm your important conversations are actually in there — don't assume.

2. Copy your character definitions manually

The export preserves data, but if a character definition is creative work you care about, save it human-readable too: copy the greeting, description, and definition fields into a text document. Exports can't always be re-imported elsewhere — plain text is your portable backup.

3. Screenshot your best conversations

For the chats that matter most — long-running stories, roleplays you'd hate to lose — screenshot or copy-paste exchanges into a document. Chats load per character (no global chat search), so go character by character, favorites first.

4. Turn off model-training use of your data

While in settings, consider the training opt-out. Mobile app: Settings → Data & Privacy → deselect "Improve the Model for Everyone." Web: Settings → Account → Manage Account & Data → Model Improvement → deselect the same. This stops new content from being used for generative-model training.

5. Don't delete your account yet — but review your subscription

Deleting is permanent and unnecessary right now: the service runs, and a buyer could keep it alive. But review your plan: annual c.ai+ subscribers with months of prepaid time have money at risk if the platform winds down. Monthly subscribers can simply cancel if things worsen.

5 Best Character.AI Alternatives (2026)

If Character.AI goes dark, the five strongest places for character chats — each verified active in 2026:

1. Janitor AI — the closest free replacement

The largest free Character.AI-style alternative with a huge community character library. Chatting on JanitorLLM is free, with optional Janitor+ for more context and priority messages.

2. Chai — best for mobile chatters

Mobile-first chat app with a massive user-made bot library and a social, swipe-and-chat feel. Freemium: free messages with limits, Premium from ~$13.99/month.

3. Talkie — best for voice-led character chat

Voice-driven conversations with characters — talk, not just type. Free with ads, Pro ~$7.99/month.

4. SillyTavern — best for full control (self-hosted)

Free, open-source, runs on your own hardware or server — your chats are truly yours and can't be shut down by anyone. Bring your own model (local or API). Real setup curve, but the most private and future-proof option.

5. Replika — best for one long-term companion

Strength is long-term memory, mood tracking, and a 3D avatar with voice. Free tier plus Pro from ~$19.99/month.

FAQ

Is Character.AI shutting down?
Not yet. It remains up while a court-supervised sale runs through late November or early December 2026. But the company warns that without a bid allowing continued operations, the platform will be wound down — so a shutdown is possible.

What happens to my Character.AI chats?
Nothing changes for now. If the platform winds down after the sale, anything stored only on its servers could be lost. Export now: Profile Settings → Account → Manage Account & Data → Export data.

How do I export my Character.AI chats?
Web: Profile → Profile Settings → Account → Manage Account & Data → Export data. Mobile app: Profile → settings icon → Export my data. Open the archive and verify your conversations are included.

Can I get a refund on my subscription?
Subscriptions remain active with billing unchanged, and the company promises advance notice of changes. No public refund policy for a wind-down exists — annual subscribers carry the most risk.

What's the best Character.AI alternative?
Janitor AI is the closest free replacement with the largest character library. For mobile: Chai; voice chat: Talkie; maximum privacy: self-hosted SillyTavern; one long-term companion: Replika.

Originally published on The Glitch Fixer — Character.AI Shutdown: Save Your Chats + 5 Alternatives

Muse Wonder: How to Get Access, Pricing & Fixes (2026)

The Glitch Fixer — Thu, 15 Oct 2026 10:52:24 +0000

Meta released its first true frontier AI model on October 12, 2026 — Muse Wonder — and within hours, developers hit the same confusion: where do you actually get it, how much does it cost, and why isn't it working the way the announcement implies? The launch got covered everywhere, but nobody explained the setup or the breakage.

This is the practical guide: what Muse Wonder is, the three access routes, real pricing, and fixes for the six most common setup problems developers are hitting right now.

What Is Muse Wonder?

Muse Wonder is Meta's frontier LLM series, built by Meta's Superintelligence Labs and announced October 12, 2026. Meta calls it the product of over a year of research and infrastructure work, trained at a gigawatt-scale data center called Prometheus. It follows the Muse Spark 1.3 release from April 2026, using the same pretraining and RL approach.

Two variants, built for "personal and professional agentic tasks" — long-running agents that plan work, call tools in parallel, delegate to sub-agents, and recover from errors:

Active Memory — carries important context about you across conversations and pulls it back when relevant, so agents working over days or weeks don't lose the plot.
Active Focus — keeps the agent grinding on a hard goal through blockers: trying alternative approaches, adapting to new information, checking in before consequential actions.

Meta claims "state-of-the-art" on agent, memory, and coding evaluations plus domain expertise in law, finance, and health. Treat Meta's benchmark numbers as vendor-reported until independent runs confirm them.

Muse Wonder vs Muse Wonder Full: Key Differences

The single biggest source of confusion — they are not available through the same channels:

	Muse Wonder	Muse Wonder Full
Model	Standard frontier variant	Maximum-reasoning variant for the hardest, longest jobs
Meta Model API	Available now (`muse-wonder-1.0`)	Not available — no API access at launch
Muse Code	Included	Included (subscriber plans)
Muse app	Rolling out to subscribers first	Rolling out to subscribers first
Reported pricing	$4 / 1M input, $20 / 1M output, $0.20 cached input	No per-token price; subscription-only
AI Index (Artificial Analysis v4.3)	52	56 (2nd globally, behind Claude Opus 5.5 at 58)

Key takeaway: if you want to call the model from your own code, only the standard Muse Wonder is available. Wonder Full — the variant with the headline benchmark scores — is locked behind a Muse Code subscription.

How to Get Muse Wonder Access: 3 Routes

Route 1: The Meta Model API (for developers)

Fastest route for your own applications. The Meta Model API is OpenAI-compatible — most existing code works if you change the base URL and model name:

Go to dev.meta.ai and create an account.
Generate an API key under API keys.
Set up billing — no free tier, so a card is required.
Point your client at https://api.meta.ai/v1 with model muse-wonder-1.0.

Minimal working call:

from openai import OpenAI
client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key="YOUR_META_API_KEY",
)
response = client.chat.completions.create(
    model="muse-wonder-1.0",
    messages=[{"role": "user", "content": "Summarize this lease and flag anything unusual."}],
)
print(response.choices[0].message.content)

The API also accepts the Anthropic Messages format (Claude-built tools work too), and the model is listed on OpenRouter as meta/muse-wonder.

Route 2: Muse Code (for coding — the only way to get Wonder Full)

Muse Code is Meta's coding agent, and currently the only way to access Wonder Full. Subscribe to a Muse Code plan (reported $50 and $200/month tiers) and Wonder Full runs as the default engine.

Route 3: The Muse App (coming — subscribers first)

Meta says Muse Wonder "will soon power experiences in Muse," rolling out to paying subscribers first ($20 or $100 monthly plans reported). No date announced. If you're waiting to try it casually, this is the queue — but don't expect it in your app today.

Muse Wonder Pricing Explained

Meta hasn't published an official pricing page — figures below are reported across multiple outlets, not Meta-confirmed:

API (Muse Wonder): ~$4 per million input tokens, ~$20 per million output tokens, ~$0.20 per million cached input tokens. Roughly equivalent to Claude Opus 5.5's headline rate — though one user reported a single Wonder Full run consuming 38% of a weekly Muse Code allowance, so long agent jobs are compute-heavy.
Muse Code (Wonder Full): reported $50/month and $200/month tiers.
Muse app: reported $20 or $100 monthly plans when the rollout begins.

Budget rule of thumb: per-token pricing looks modest, but agentic workloads burn tokens fast — especially reasoning models on long tasks. Set billing alerts before running agents unattended.

6 Common Muse Wonder Problems & Fixes

1. "Muse Wonder is not available in my country"

API and Muse Code availability is region-dependent. Fix: check Meta's developer docs for the supported-regions list. This is account-region enforcement — a VPN won't reliably fix billed API access.

2. 400 error when setting `reasoning_effort` to "none"

reasoning_effort accepts minimal through max — it does not support none, and sending it returns a 400. Fix: omit reasoning_effort entirely (the model picks a level) or use minimal.

3. "I can't find Wonder Full on the API"

By design, not a bug: Wonder Full is not served through the Meta Model API at launch. No hidden endpoint exists. Fix: subscribe to a Muse Code plan — the only announced route to the Full variant.

4. Reasoning doesn't carry over between turns

On the Chat Completions endpoint, reasoning doesn't persist across turns, making multi-turn agent conversations inconsistent. Fix: use the Responses API with previous_response_id to preserve reasoning context.

5. "Invalid API key" or 401 errors on first call

No free tier — billing must be set up on your Meta developer account or calls fail at auth. Fix: confirm the key was generated under API keys at dev.meta.ai, billing is active, and you're using https://api.meta.ai/v1 as the base URL (not a legacy Meta AI endpoint).

6. Muse app doesn't show Muse Wonder yet

The rollout is announced as forthcoming for subscribers — no date. Fix: nothing to toggle. Paid plans ($20/$100 reported) are first in line; use the API route (Route 1) meanwhile.

Benchmark Snapshot: Wonder Full vs Claude Opus 5.5

Launch-reported figures (vendor-reported where noted — independent confirmation pending):

Benchmark	Muse Wonder Full	Claude Opus 5.5
AI Intelligence Index v4.3	56 (2nd)	58
DeepSWE 1.1 (coding)	79%	74.2%
SWE-Marathon (long tasks)	57.5% (Meta)	46.3% (Meta)
MemoryBench (long-term recall)	Leads (Meta)	Behind

Pattern: Wonder Full leads on long-horizon coding evals but trails Opus 5.5 on the overall index. Meta's safety-evaluation report wasn't published at launch — worth watching before production workloads.

FAQ

How do I get access to Muse Wonder?
Three routes: (1) the Meta Model API at dev.meta.ai (muse-wonder-1.0, OpenAI-compatible, billing required), (2) a Muse Code subscription — the only way to get Wonder Full, (3) the Muse app, rolling out to subscribers first, no date announced.

How much does Muse Wonder cost?
Reported (unconfirmed) API pricing: $4/1M input, $20/1M output, $0.20/1M cached input tokens. Muse Code plans with Wonder Full reported at $50 and $200/month. No official pricing page yet.

What's the difference between Muse Wonder and Wonder Full?
Standard Wonder is available via API and Muse Code. Wonder Full is the maximum-reasoning variant, available only to Muse Code subscribers — not through the API at launch.

Can I use Muse Wonder for free?
No. No free or contributor tier — the API needs billing set up, and Muse Code/app access both require paid subscriptions.

Originally published on The Glitch Fixer — Muse Wonder: How to Get Access, Pricing & Fixes (2026)

OpenAI Dots: How to Set Up & Use ChatGPT's Always-On AI Agent (2026)

The Glitch Fixer — Thu, 15 Oct 2026 10:51:40 +0000

On 29 September 2026, OpenAI announced Dots at DevDay — persistent AI agents inside ChatGPT that keep working on your goals between conversations. A Dot isn't a longer chat session: each one gets its own cloud computer and browser, runs on GPT-6 Astra, connects to 4,000+ apps via plugins, and can research, draft, code, and report back while you're away. You reach the same Dot through ChatGPT on desktop and mobile, through Slack and Microsoft Teams, or even by voice call.

The catch: Dots aren't available to everyone, they can't be created on a phone, and plenty of people on eligible plans still can't see them. This guide covers what Dots actually are, who qualifies, the exact setup steps, how the safety controls work, and what to do if Dots are missing from your account.

What Are OpenAI Dots?

An OpenAI Dot is a persistent personal AI agent inside ChatGPT. The difference from ordinary ChatGPT is persistence plus its own machine. Instead of asking ChatGPT to "analyse this month's customer feedback" and getting one answer, you give a Dot ongoing responsibility for the job: it checks connected apps in the background, tracks how the data changes, and reports back with progress.

Each Dot has its own cloud computer, browser, and memory. You give it a goal, connect the apps it needs, set the boundaries in what OpenAI calls Custom Rules, and it keeps making progress between conversations. It can contact you with updates, ask questions when unsure, and hand back decisions needing human judgement. You can open its cloud computer at any time to see what it's doing, and in the desktop app an Activity View shows the steps it has delegated.

Practical examples from OpenAI: keeping a sales proposal updated as requirements change, investigating a newly reported bug when it appears, preparing content as new source material arrives.

Who Can Get OpenAI Dots (and What Do They Cost)?

Dots are still rolling out and limited to paid tiers. Eligibility as of October 2026:

ChatGPT Pro 100 / 200 / 500 ($100 / $200 / $500 per month) — adults only, excluding the EEA, Switzerland, and the UK. Your first Dot is included at no extra cost.
ChatGPT Business Premium ($100 per user per month billed annually, or $125 monthly) — all supported regions.
Enterprise, Edu, and Healthcare — Dots are in beta and off by default; a workspace admin must enable them.
Free, Go ($8), and Plus ($20) plans — not included. OpenAI says it plans to expand access later, with no date.

Talking to your Dot doesn't count against ChatGPT message limits, but the work it starts does: tasks it hands to ChatGPT Work or Codex count as usual. OpenAI has said users will eventually be able to add more Dots or capacity, but pricing hasn't been published.

Rollout is gradual — Dots "may take several days to reach your account" even on an eligible plan. If you qualify and still can't see them, wait a few days before assuming something is broken.

How to Set Up Your First OpenAI Dot (Step by Step)

Dots are created on desktop only — you can't create one on a phone or mobile web. Once it exists, you can message it from the ChatGPT mobile app.

Open the Dots section on desktop. Use the ChatGPT desktop app for macOS or Windows, or go to chatgpt.com/dots in a desktop browser, signed in to an eligible account.
Create your Dot and name it. The name becomes its handle, in the form @yourname-name (e.g. name it "Alfred" → @yourname-alfred). You'll use this handle to mention it in Slack, Teams, and ChatGPT.
Connect your apps in the Plugins tab. Link the apps your Dot needs — email, calendar, docs, project tools. Permissions are shared across ChatGPT, Work, and Codex, so review each connection before approving.
Set your Custom Rules. The most important step: for each Dot you decide what it can do on its own, what needs pre-approval, what it must always ask about first, and what's off-limits. Read the safety section below before skimming this.
Leave local-computer access off unless you need it. Dots can access your own computer, but this is off by default. Only enable it for a specific reason, then turn it back off.
Give it its first job. Start with something bounded and reviewable — "watch this shared document and tell me when the client comments," or "summarise new project emails every morning." Track progress in the Dot's profile: In progress, Scheduled, Completed lists.

After setup, reach your Dot anywhere: message it in ChatGPT (desktop or mobile), tag its handle in Slack or Teams, or call it by voice. Texting support is currently limited to a small Pro beta in the US.

Set Custom Rules Before You Connect Everything

A Dot's background research is read-only by default — it can't send messages or change anything in your apps while idly watching. Anything with outside effects goes through rules you define:

Take action — the Dot acts on its own.
Act if pre-approved — it acts only on pre-approved items.
Ask first — it must check with you every time.
Hand off — it stops and gives the step back to you.

OpenAI also hard-codes two guardrails: password changes and money movement always stay human, and a separate Auto-review system examines planned actions before they run. On prompt injection, OpenAI reports no scored successes in 16,600 generated attack emails — but scope problems grow as task chains get longer. Practical takeaway: start conservative (Ask first on anything consequential) and review consequential work, especially early on.

Why Don't I See Dots in My ChatGPT Account?

The most common Dots problem right now — the fix depends on which applies:

You're on Free, Go ($8), or Plus ($20). Dots aren't included on these plans at launch. No setting will surface them.
You're in the EEA, UK, or Switzerland on a Pro plan. Dots are excluded for Pro users in these regions. Business Premium users in all supported regions are eligible.
Your account is eligible but the rollout hasn't reached you. OpenAI confirms this can take several days. Check back periodically; nothing to fix on your end.
You're on mobile. Dots can't be created from the mobile app or mobile web. Use a desktop browser or the desktop app.
Your Enterprise/Edu/Healthcare workspace hasn't enabled it. Off by default — ask your workspace admin to enable the beta.
You're under 18. Dots require an adult account.

FAQ

What is OpenAI Dots?
Always-on AI agents in ChatGPT, announced at DevDay on 29 September 2026. Each Dot runs on GPT-6 Astra with its own cloud computer and browser, connects to 4,000+ apps, and keeps working toward your goals between conversations — reporting back through ChatGPT, voice, Slack, or Microsoft Teams.

Which ChatGPT plans include Dots?
Pro 100/200/500 (adults, excluding the EEA, Switzerland, and the UK), Business Premium (all supported regions), and Enterprise/Edu/Healthcare where an admin enables the beta. Free, Go, and Plus don't include Dots at launch.

How much do OpenAI Dots cost?
Your first Dot is included with an eligible plan. Pro costs $100, $200, or $500/month; Business Premium is $100 per user/month billed annually ($125 monthly). Pricing for additional Dots hasn't been published.

Can I create a Dot on my phone?
No. Creation requires the ChatGPT desktop app (macOS/Windows) or desktop web. After creation you can message your Dot from the mobile app.

Is Dots available in Europe and the UK?
Not for Pro users — excluded in the EEA, Switzerland, and the UK at launch. Business Premium users in all supported regions are eligible.

Is ChatGPT Agent the same as OpenAI Dots?
No. The old Agent mode is discontinued; Dots are persistent assistants that can start Work and Codex tasks and keep working on ongoing goals between conversations.

Originally published on The Glitch Fixer — OpenAI Dots: How to Set Up & Use ChatGPT's AI Agent (2026)