The Claude Code extension backdoor came from a stolen vendor access key, Cursor confirmed on 13 November 2026. Tampered builds of Anthropic’s extension, versions 2.1.312 and 2.1.313, reached as many as 5,000 users during an 11-hour window on 29 October. Updating the extension does not remove what it left behind.

What Cursor Confirmed About the Breach

Cursor, the San Francisco company behind the AI coding editor of the same name, published the results of its investigation on 13 November 2026, about two weeks after it first disclosed the incident.

The company traced the attack to Fireworks, a third-party AI inference provider, which exposed an access key that Cursor had issued to it. Cursor’s account describes a chain of four steps:

  • The key: The attacker used the exposed key to copy data from a read-only storage bucket that Cursor shared with the vendor.
  • The credentials: Some of the stored coding-agent traces in that bucket contained credentials that had not been redacted.
  • The spread: Those credentials gave the attacker access to several software-as-a-service tools that Cursor uses internally.
  • The extension: One stolen credential was used to place malicious copies of the Claude Code extension into Cursor’s marketplace cache.

Cursor’s notice does not say how Fireworks exposed the key, and it does not disclose the key itself. The company says the last attacker activity it observed was on 1 November 2026, and that it was blocked. Its report states: “We believe the incident is contained.”

Cursor was also explicit that Anthropic’s own systems were not involved. “Anthropic did not publish this copy, and this was not a vulnerability in Anthropic’s product,” the company wrote. Felix Rieseberg of Anthropic told CSO Online that “the versions on our official channels are not affected, and this only impacts Cursor users.”

Who Received the Tampered Build, and Who Did Not

Only one narrow group was exposed: Cursor users on Windows or macOS who installed or updated the Claude Code extension from Cursor’s own marketplace during a single window on 29 October.

In scopeNot in scope
Anthropic.claude-code versions 2.1.312 and 2.1.313All other versions of the extension
Windows and macOSAnthropic’s official distribution channels
The Cursor VS Code Extension MarketplaceOpen VSX, Microsoft’s Visual Studio Marketplace and the npm package
Roughly 2:27am to 1:35pm Pacific Time on 29 October 2026Other extensions in Cursor’s marketplace, which Cursor says it validated as authentic
At most 5,000 users who requested those cached versionsCursor’s production systems, its app, official updates and release signing keys

A separate set of data was taken earlier in the same intrusion, before the extension was touched. Cursor says fewer than 0.01% of customers had API credentials exposed, and fewer than 0.05% of users without Privacy Mode had coding-agent traces exposed, most of them from 2025. The attacker also reached archived non-production source-code repositories, limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding-agent conversations, and credentials that users had sent to Cursor’s support team. Prompts and code belonging to Privacy Mode users, and customer code repositories, were not affected.

The figures Cursor published in its incident report are the company’s own estimates; no independent audit of the scope has been published.

How to Check Your Machine for the Attacker’s Key

The single reliable indicator is an unauthorised SSH public key. The malicious code appended an ssh-ed25519 key carrying the comment svc-telemetry, with the fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk.

  1. Check the key file: On macOS and Linux, search ~/.ssh/authorized_keys for the text svc-telemetry. On Windows, check both %USERPROFILE%\.ssh\authorized_keys and C:\ProgramData\ssh\administrators_authorized_keys.
  2. Confirm the fingerprint: Match the key against the fingerprint above rather than trusting the comment, which an attacker could change.
  3. Preserve evidence first: Microsoft advises keeping the extension package, the modified SSH files, the generated keys and scripts, and the startup entries before deleting anything.
  4. Remove only that entry: Delete the attacker’s key line, leaving your own keys in place.
  5. Rotate what the host could reach: Change credentials and revoke sessions from a device you know is clean.

Cursor published remediation scripts for two platforms, remediate-macos.sh and remediate-windows.ps1, which remove the extension, the persistence entries, the tunnel processes and the attacker’s key. The company says the scripts exit with a non-zero status if anything is left behind. Its notice lists Linux indicators but does not offer a Linux script.

A clean result is a negative on one indicator, not a clearance for the whole machine. Cursor’s response steps ask administrators to search endpoint, network and extension-installation telemetry for the published indicators, and to isolate any host showing evidence that the payload ran.

Why Removing the Extension Does Not Clean the Machine

Uninstalling or updating the extension leaves the backdoor’s scaffolding in place, because the payload wrote files outside the extension’s own directory. Cursor states plainly that “uninstalling the extension may not be sufficient,” because the payload “drops standalone persistence scripts that keep running.”

PlatformPersistence mechanismFiles to look for
macOSUser LaunchAgent~/Library/LaunchAgents/com.apple.sysmond.plist, ~/.config/.smt, ~/.config/.smtk
LinuxCrontab @reboot entry~/.config/.smt, ~/.config/.smtk, /tmp/.stk_ plus ten hex characters
WindowsScheduled task named SysMon%USERPROFILE%\.config\smt.cmd, %USERPROFILE%\.config\smtk
Windows fallbackStartup-folder batch filesysmon.cmd in the Start Menu Startup folder

The names were chosen to blend in. SysMon and com.apple.sysmond both resemble legitimate system-monitoring components, and Microsoft’s analysts warn administrators to read each object’s configured action before deleting it, in case a genuine tool shares the name. The scripts relaunched the connection roughly 60 seconds after each failure.

The outbound destination was the address 134.209.82.225, used both for a reverse SSH tunnel on TCP port 22 and for a beacon on port 8099. Cursor worked with the hosting provider to take that server offline.

Steve Povolny of Exabeam put the problem to CSO Online this way: “removing the extension only sends the visitor away while the spare key stays on the hook.”

What the Code Collected, and What to Rotate

The malicious loader ran as soon as the extension module loaded, before the extension’s normal code, so a user did not have to open Claude Code or send a prompt for it to fire.

According to Microsoft Threat Intelligence, it gathered the hostname, username, operating system, architecture, home directory, working directory and the globally configured Git email address. It then looked for environment variables whose names contained strings such as KEY, TOKEN, SECRET, PASS, AWS or GCP, keeping up to about forty matches. Microsoft was specific that the code collected the variable names, not their values.

That distinction matters for how far a response needs to go. Variable names tell an attacker which secrets exist on a machine and what to come back for; they do not hand over the secrets themselves. Both Cursor and Microsoft still advise rotating credentials and sessions that were reachable from an affected host, and doing it from a known-clean device.

Developers who have dealt with the recent run of attacks on AI coding tools will recognise the pattern. It is the same reasoning behind the AGENTS.md attack, which used a developer’s own coding agent to run instructions planted on their machine, and behind the ConnectWise Cloud Backup breach, where a trusted integration became the way in.

Why the Backdoor Failed on Most Machines

The attack was noisier than it was successful. Cursor reported that errors in the malicious code, combined with the administrator permissions that enabling SSH requires, blocked SSH access on most macOS and Windows hosts.

On macOS, the loader checked whether Remote Login was switched off using a command that needs administrator rights; without them the command returned an access error and the loader skipped the step. On Windows, the security firm Phorion found a flaw that could stop the command setting up the SSH service from running correctly.

Two caveats keep this from being reassuring. The loader could still append its key, write its persistence files and send the host details even when the tunnel never opened. And a running SSH process on an affected machine does not by itself prove the attacker got in, nor does its absence prove they did not.

Where the Incident Stands as of 16 November 2026

As of 16 November 2026, Cursor says the incident is contained, the marketplace has been rebuilt and restored, and no attacker activity has been observed since 1 November. The company moved all inference traffic off Fireworks, deactivated the vendor’s key, cut its access to shared datasets, rotated exposed credentials, redeployed its production infrastructure, added secret scanning across its code repositories, and brought in outside forensics specialists.

Three things remain unresolved. Fireworks has not published a technical account of how the access key was exposed, so the attribution of that exposure rests on Cursor’s account. No public report confirms that any developer machine was actually accessed through the tunnel. And Cursor’s notice does not name a version of the extension as safe, although security trade press has reported that Anthropic’s legitimate extension moved to version 2.1.314 — a detail worth treating as reported rather than confirmed.

The same vendor breach has surfaced elsewhere. Fireworks was also the route into the Uber driver data breach disclosed on 13 November, which affected thousands of US drivers.

Frequently Asked Questions

Was I at Risk if I Use Claude Code in VS Code?

No. Cursor says it has no evidence that other versions of the extension or the official distribution were affected, and Microsoft confirmed that the Visual Studio Marketplace and npm packages were not. Only installs from Cursor’s own marketplace during the 29 October window are in scope.

Does Updating the Extension Remove the Claude Code Extension Backdoor?

No. The payload wrote startup entries and scripts outside the extension folder, and those keep running after the extension is replaced. The attacker’s SSH key has to be found and deleted separately.

What Should I Search For to Know if My Machine Was Touched?

Look in your authorized_keys file for an ssh-ed25519 key commented svc-telemetry, then verify it against the fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk. Also check for the LaunchAgent, scheduled task or crontab entry listed above.

Was Anthropic Breached?

No. Cursor stated that Anthropic did not publish the tampered copy and that the incident was not a vulnerability in Anthropic’s product. The attacker reached Cursor’s marketplace, not Anthropic’s systems.

Do I Need to Rebuild the Computer?

Cursor’s response steps leave that to the administrator’s judgement, asking them to decide whether a host showing evidence of the backdoor should be rebuilt. Microsoft advises rebuilding where device trust cannot be restored, and verifying afterwards that the artefacts do not return after a restart or login.