Cursor's Nov 13 postmortem traces the tampered Claude Code extension to a key exposed at Fireworks. Who was hit, what to check, and what helps.
If you installed or updated the Claude Code extension from Cursor's extension marketplace on macOS or Windows on October 29, 2026, between about 2:27 a.m. and 1:35 p.m. PT, you may have received a tampered build, version 2.1.312 or 2.1.313, that tries to plant an SSH backdoor. Microsoft says removing the extension does not remove what it wrote outside the extension folder. Cursor's November 13 postmortem now explains how the attacker got that far: an access key that Fireworks exposed led, in several steps, to tampered copies placed directly on the cache that serves Cursor's marketplace.
Sources: Cursor: An update on our recent security incident (Nov 13, 2026), Cursor: Security notice for Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026), Microsoft Security Research write-up (Oct 30, 2026)
Disclosure: we build TLM Forge, a process layer for Claude Code, so weigh the product section accordingly. That section is short, and it says plainly that TLM Forge does not detect or block an attack like this one.
Who was affected, and who was not
Cursor's security notice, published October 29 and revised the next day, names Anthropic.claude-code versions 2.1.312 and 2.1.313 for Windows and macOS. The postmortem says "At most 5,000 users who requested specific cached versions of the extension on October 29 potentially received the malicious build," and that it "was available for approximately 11 hours before we removed it." The 5,000 figure is a ceiling, not a count of infected machines.
Sources: Cursor: Security notice for Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026), Cursor: An update on our recent security incident (Nov 13, 2026)
Anthropic's position, as The Register quotes it, is that a malicious actor used access to Cursor's marketplace to publish a tampered copy that Anthropic did not build or publish, and that its official channels are unaffected. Microsoft confirmed the package was never published to the Visual Studio Marketplace or through the @anthropic-ai/claude-code npm package. Cursor adds that the marketplace source, Open VSX, never had the malware.
Sources: The Register (Oct 30, 2026), Microsoft Security Research write-up (Oct 30, 2026), Cursor: An update on our recent security incident (Nov 13, 2026)
- Exposed: Cursor on macOS or Windows, with the extension installed or auto-updated from Cursor's marketplace inside the October 29 window.
- Not affected, per the vendors: the Claude Code CLI from the official npm package, the extension from the Visual Studio Marketplace, and the extension from Open VSX.
- Not settled by version alone: StepSecurity downloaded the 2.1.312 build that Open VSX served and verified it clean, so the same version number can be safe or tampered depending on where it came from.
What early coverage got wrong or left open
The first days of coverage produced a run of corrections. Several first reports named Open VSX as the source, and The Register reported that prominent security figures had alleged a breach at Anthropic, which Anthropic denied. StepSecurity, Phorion, OpenSourceMalware and The Hacker News each added a correction or update saying the tampered build came from Cursor's marketplace. On October 30, The Register reported that neither Cursor nor Anthropic had said how the package reached that marketplace, and The Hacker News listed the method as an open question. The table sets the early picture against what the primary sources say by November 13.
Sources: Help Net Security (Oct 30, 2026), StepSecurity analysis (Oct 30, 2026, updated through Nov 1), Phorion advisory (Oct 30, 2026, updated Oct 31), OpenSourceMalware analysis (Oct 30, 2026), The Hacker News (Oct 30, 2026), The Register (Oct 30, 2026)
| Question | Early coverage (October 30) | Primary sources by November 13 |
|---|---|---|
| Which marketplace carried it? | Several first reports named Open VSX | Cursor: Open VSX never had the malware. The copies sat on the cache that serves Cursor's marketplace |
| Which versions? | 2.1.313 only | Cursor revised its notice to 2.1.312 and 2.1.313, for Windows and macOS |
| Was Anthropic breached? | Alleged by some security figures, per The Register | Anthropic: not a compromise of its systems. Cursor: not a vulnerability in Anthropic's product |
| How did the attacker get in? | Unexplained; neither company had said | An access key Cursor issued to Fireworks was exposed by Fireworks |
| Was the extension the whole incident? | Cursor's first notice described an incident involving one specific extension | Cursor: no. The attacker also copied agent traces and reached SaaS tools, archived non-production source repositories and support data |
How the attack worked, per Cursor's postmortem
Cursor's account runs in a straight line. An attacker obtained an access key that Cursor had issued to Fireworks, which Cursor says Fireworks exposed. The key opened a read-only storage bucket that Cursor used to share data with Fireworks. That bucket held "limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which we were permitted to use for model training."
Sources: Cursor: An update on our recent security incident (Nov 13, 2026)
A small number of those traces contained credentials nobody had redacted. Cursor says those credentials let the attacker reach select SaaS tools, and through them archived non-production source repositories, limited support data, product usage metrics, summaries of some coding agent conversations and credentials that users had sent to support. On October 29, a stolen credential from that exposed data let the attacker place the tampered extension copies on the marketplace cache. Cursor's figures: for users not using Privacy Mode, "coding agent traces were exposed for fewer than 0.05% of users," and "less than 0.01% of customers had API credentials exposed." The last attacker activity Cursor saw was on November 1, and it was blocked.
Sources: Cursor: An update on our recent security incident (Nov 13, 2026)
Fireworks published two bulletins about its own October incidents. The first, published October 9, describes unauthorized use of internal Fireworks credentials and access to environment variables used in evaluation jobs in a limited number of customer accounts. The second, last updated November 9, adds account and billing information and, for a limited set of accounts, secrets, datasets and incomplete models. It also says a third-party cloud provider that hosted Fireworks R&D machines had an incident on October 23, and it says the evidence indicates links to a financially motivated group that sells tokens from unauthorized LLM access, a practice it calls LLMjacking. Neither bulletin mentions Cursor, so they do not say which Fireworks incident exposed Cursor's key. Cursor's postmortem does not attribute the attack to any group.
Sources: Fireworks AI: October 2026 security incident bulletin, Fireworks AI: Update on October 2026 security incident (Nov 9, 2026)
Wiz counted at least nine AI developer platforms that disclosed incidents in October, Fireworks AI first, on October 6, and wrote that no single vulnerability or dependency links the attacks. Its post does not mention Cursor. The Cursor incident came in a month when many AI developer platforms disclosed incidents. Cursor says it was not a vulnerability in Anthropic's product.
Sources: Wiz: "Black October" for AI Developer Platforms
Why one tampered file was enough
An extension is code that runs with your privileges. VS Code's documentation says the extension host "has the same permissions as VS Code itself," so an extension can read and write files, make network requests and run external processes. Phorion makes the same point for Cursor: extensions run with the permissions of the IDE and the user running it. StepSecurity's analysis adds that the payload runs under plain Node.js, that "loading the extension module is enough," and that no Claude conversation or package-manager hook is involved.
Sources: VS Code docs: Extension runtime security, Phorion advisory (Oct 30, 2026, updated Oct 31), StepSecurity analysis (Oct 30, 2026, updated through Nov 1)
The tampered copy also looked routine. The Hacker News reports it shipped under the extension's legitimate publisher identity, anthropic.claude-code. Claude Code's docs say the extension's version number is the Claude Code version it bundles, so 2.1.313 read like any other point release. StepSecurity found that the bundle inside the 2.1.313 package still identified itself as 2.1.312, and that the file matched the clean 2.1.312 bundle it downloaded apart from a single 7,386-byte loader added in front. Ax Sharma of Manifold Security told Help Net Security: "Checking the publisher name doesn't help when the publisher is real."
Sources: The Hacker News (Oct 30, 2026), Claude Code docs: Use Claude Code in VS Code, StepSecurity analysis (Oct 30, 2026, updated through Nov 1), Help Net Security (Oct 30, 2026)
The loader failed quietly, too. StepSecurity says each of its six stages sits in a try/catch that records a status string instead of failing. One stage did break on most machines: Cursor says turning on the local SSH service requires admin rights, which the extension lacks on most macOS and Windows hosts, so "the attacker could not have logged in through the tunnel" there. Even so, "The extension added the attacker's SSH key, installed persistence, and sent host details to the attacker." Those details included the names, not the values, of up to 40 environment variables matching patterns such as KEY, TOKEN and SECRET.
Sources: StepSecurity analysis (Oct 30, 2026, updated through Nov 1), Cursor: Security notice for Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026), The Hacker News (Oct 30, 2026)
Check your machine
Start with exposure, because the version number alone does not settle it. Open the extension in the Extensions view and read its version. Then ask whether it came from Cursor's marketplace on macOS or Windows inside the October 29 window, by fresh install or by auto-update. Jenn Gile, co-founder of OpenSourceMalware, put the auto-update point this way: "It was pulled very quickly, but anyone who has their Claude Code instance set to auto-update could have consumed the malware."
Sources: The Hacker News (Oct 30, 2026), Cursor: Security notice for Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026)
Cursor publishes its own check and remediation scripts for macOS and Windows in the October 29 notice, so run those first. The read-only commands below check some of the indicators Cursor lists for macOS and Linux, and they skip the temporary .stk_ key file. They were checked on macOS (Darwin 25.5.0) against throwaway home directories, with and without the indicator files, and they print nothing when nothing matches.
Sources: Cursor: Security notice for Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, 2026)
grep -ns 'svc-telemetry' ~/.ssh/authorized_keysprints the line if the attacker's key label is present.ls -la ~/.config/.smt ~/.config/.smtk ~/Library/LaunchAgents/com.apple.sysmond.plist 2>/dev/nulllists the persistence files named in Cursor's indicators.pgrep -fl 'ssh -N -R|autossh -M 0'shows a running reverse tunnel.crontab -l 2>/dev/null | grep -n '\.config/\.smt'finds the reboot entry the loader adds on Linux.
A clean result means none of the published indicators were found. It does not prove the machine is clean. If you installed an affected build, follow Cursor's remediation steps anyway, and rotate credentials from a trusted system.
If you were hit: what Cursor and Microsoft say to do
- Isolate any host showing evidence that the payload ran or that the backdoor was installed.
- Remove the unauthorized SSH key, the persistence entries and any running tunnel processes, and decide whether to rebuild the host. Cursor offers remediation scripts for macOS and Windows.
- Remove those artifacts separately from the extension. Microsoft says removing or updating the extension does not remove artifacts written outside its installation directory, and Cursor warns that uninstalling alone may not be enough.
- Rotate credentials and sessions reachable from the machine. The loader sent variable names, but researcher Paul McCarty noted that "successful SSH access would allow the operator to inspect the account directly."
- If your team admin received an email from Cursor listing specific keys to rotate, rotate those. Cursor says it also asked the services that issued the credentials to revoke them.
What would have helped, control by control
Microsoft's write-up recommends that organizations using Cursor restrict extensions to approved publishers and configure an extension installation cooldown. Both are real controls. Whether either would have stopped this incident is a separate question, and the table answers it from the documentation. None of it was tested against the tampered builds.
Sources: Microsoft Security Research write-up (Oct 30, 2026), Cursor docs: Extensions
| Control | What the docs say it does | Fit for this incident |
|---|---|---|
| Verified publisher badge | VS Code docs: the publisher proved domain ownership. Cursor docs: verification confirms publisher identity | Weak. The Hacker News reports the copy shipped under the legitimate publisher identity, so a check on who published it had nothing to flag |
| Allowlist of publishers or IDs | Cursor docs: restrict installs to approved publishers or extension IDs, through the team dashboard or MDM | Weak. A team that allows Anthropic's Claude Code extension would have allowed a copy that used the same ID |
| Marketplace malware analysis | Cursor docs: automated malware and supply-chain analysis before an extension is shown in search or served | Unclear. Cursor says the copies went directly onto the serving cache; the postmortem does not say whether analysis ran on them |
| Install cooldown or update delay | Cursor docs: admins can defer installs and updates until a version has been published for a set number of hours, default 0 (off). VS Code docs: autoUpdateDelay, default 12 hours | Possible. The exposure window was about 11 hours. The docs do not say how publish time is measured for a file placed on a cache |
| Signature verification | Cursor docs: enterprise teams can require valid Open VSX signatures before install | Unknown. The postmortem does not say whether the tampered files were signed |
| Editor without admin rights | Cursor notice: turning on SSH needs admin rights, which the extension lacked on most hosts | Partial. Per Cursor, errors in the malicious code made that step fail on most hosts, so the attacker could not have logged in through the tunnel. The key, persistence and host report still ran |
| Endpoint detection | Microsoft lists Defender alerts such as "Suspicious addition of an SSH key" and "Suspicious SSH tunneling activity" | After the fact. It flags what the payload does, not the bad update |
Read the table in two halves. Identity checks and allowlists answer who published an extension, and this incident was about what a trusted ID was serving. Delay-based controls and endpoint detection do not depend on who published the file. A cooldown buys time for a bad version to be pulled, and endpoint detection flags what the payload does. In Cursor, a cooldown is a setting rather than a new tool. It has a cost: a delay also holds back legitimate fixes. Cursor's docs add that cooldown and signature checks are client-side install controls that complement, but do not replace, marketplace-layer analysis and blocking.
Sources: Cursor docs: Extensions, VS Code docs: Extension Marketplace (auto-update)
The part that applies to any coding agent: transcripts hold secrets
One link in Cursor's chain was a store of session traces that held a few unredacted credentials. The same pattern exists on your own machine. Claude Code's data-usage documentation says clients store session transcripts locally in plaintext under ~/.claude/projects/ for 30 days by default, adjustable with cleanupPeriodDays, and the .claude directory reference gives the minimum as 1. Treat anything that appeared in a session, such as command output that printed an environment variable, as sitting in plaintext on disk. A compromised host gives an attacker the same read access to those files that you have.
Sources: Claude Code docs: Data usage, Claude Code docs: The .claude directory
- Keep secrets out of the conversation in the first place, and rotate on exposure. AI Agent Tokens Need Rotation, Not Just Scope covers why agent tokens need rotation as well as narrow scope.
- If you never resume old sessions, shorten the retention period with cleanupPeriodDays to shrink what a stolen disk or a compromised host can read.
- If you use Cursor, check that Privacy Mode is on. Cursor lists "Your prompts and code if you use Privacy Mode" among what was not affected.
Extensions are one of several code paths that run as you
A verified publisher badge says who published an extension, not what a file contains. For servers that an agent connects to, see How to Vet MCP Servers Before You Trust Them. For another extension path, see Claude Code Mods Ship Unsandboxed, On by Default. For a malicious package that rewrote agent config files, see Tensorlake npm Worm Rewrote Your AI Agent Config Files.
Sources: Cursor docs: Extensions
Where TLM Forge fits, and where it does not
TLM Forge does not detect or block this kind of attack. It works inside a Claude Code session: it sizes the task as Light, Medium or Deep, shows a design for your OK on Medium and Deep work, runs independent reviewer agents over the final diff on Medium and Deep work, and uses a commit hook that blocks git commit, git push and gh pr merge when the code changed since the final review. The tampered extension ran when the editor loaded it, before any session or design step existed. The hook matches the text of those three commands and sees only commands Claude Code runs, so it sees nothing an attacker runs from another shell.
The response to an incident like this is the checklist above: check exposure, remove the artifacts, rotate what was reachable, and consider a delay on updates. TLM Forge is not part of that list.
Frequently asked questions
01Is the Claude Code extension safe to use after the Cursor incident?
The tampered builds, 2.1.312 and 2.1.313 served from Cursor's marketplace on October 29, were removed, and Cursor says it reviewed and rebuilt the marketplace. Anthropic says its own channels were not affected. If you installed either version from Cursor's marketplace in the exposure window, run Cursor's check script first.
02How do I know if I installed the malicious Claude Code extension in Cursor?
Cursor scopes exposure to version 2.1.312 or 2.1.313 obtained from Cursor's marketplace on macOS or Windows between about 2:27 a.m. and 1:35 p.m. PT on October 29, 2026. Its notice says to look for an SSH authorized key labeled svc-telemetry and provides check scripts.
03Does uninstalling the extension remove the backdoor?
No. Cursor says the payload drops standalone persistence scripts that keep running. Microsoft says removing or updating the extension does not remove artifacts written outside its installation directory, such as SSH keys, scripts and startup entries. Remove those separately, using Cursor's scripts or the published indicators, and rotate credentials reachable from that machine.
04Was Anthropic hacked in the Claude Code extension incident?
Anthropic says no: it was not a compromise of its systems or employee laptops. Cursor's postmortem says an attacker used a stolen credential to place tampered copies on the cache that serves its marketplace, and calls it not a vulnerability in Anthropic's product. Open VSX never carried the malware, per Cursor.
05Can editor settings stop a bad extension update?
Partly. VS Code documents an auto-update delay of 12 hours by default, and Cursor documents an install cooldown for teams that is off by default. A delay helps only if a bad version is pulled before it expires. VS Code's delay applies only when auto-update is on, and pressing Update bypasses it.