Get a demo — 30 minutes →
← Back to blog
Anomity robot illustrating Claude Code Extension 2.1.313 SSH Backdoor - Cursor Marketplace, Not Open VSX
AdvisoryCritical

Claude Code Extension 2.1.313 SSH Backdoor - Cursor Marketplace, Not Open VSX

IDE Extensions & Plugins·Critical·Trojanized anthropic.claude-code 2.1.313, disclosed October 30, 2026 (OSV MAL-2026-18055; no CVE)·
Affected Cursor users who installed or auto-updated to anthropic.claude-code 2.1.313 from Cursor's extension marketplace between about 09:27 and 20:35 UTC on October 29, 2026. Cursor puts the reach at up to 5,000 users. The builds on the Microsoft Visual Studio Marketplace, Open VSX and npm were not affected, per Anthropic and Microsoft.

For about eleven hours on October 29, 2026, developers using Cursor who installed or updated the official Claude Code for VS Code extension received a build that was not Anthropic's. Version 2.1.313 of anthropic.claude-code carried a short loader that planted an attacker's SSH key, opened a reverse tunnel out of the machine, and installed persistence that survives removing the extension. Cursor says up to 5,000 users may have installed it. This advisory covers what is confirmed, why the early Open VSX attribution was wrong, the indicators to hunt for, and how to clean up.

The incident matters beyond its install count because of what was impersonated. This was not a lookalike package from an unknown publisher, the pattern in the Solidity Pro credential stealer. It was the real extension ID of the most widely deployed AI coding agent, delivered through an editor's built-in update path. Every review heuristic that asks "is this the publisher we expect?" passed.

What happened

Update, November 13, 2026: Cursor has published how the attacker got in. It traces the incident to an access key it had issued to Fireworks AI, which was exposed in a breach at Fireworks and used to copy coding agent traces holding some unredacted credentials. A stolen credential from that data let the attacker place the tampered build on the cache behind Cursor's marketplace. Cursor's update asks anyone who installed 2.1.312 or 2.1.313 from its marketplace during the window to follow its notice. We cover the full chain in how the Fireworks AI breach led to Cursor's malicious extension.

Cursor runs its own extension marketplace for its VS Code-based editor. On October 29, a modified build labelled anthropic.claude-code 2.1.313 appeared there under the legitimate extension identity, rather than as a separate fake listing. Per Cursor's figures, it was available from about 2:27 a.m. to 1:35 p.m. Pacific (09:27 to 20:35 UTC). Researchers at Phorion and OpenSourceMalware published the first analyses on October 30, and Microsoft published its own write-up the same day.

Anthropic's position, stated by members of its technical staff, is that this was not a compromise of Anthropic's systems or employee laptops: a malicious actor used access to Cursor's marketplace to publish a tampered copy that Anthropic did not build or publish. Anthropic confirmed 2.1.313 was never published to Open VSX, the Microsoft Visual Studio Marketplace, or any of its official hosts, and skipped to 2.1.314 to avoid confusion. Microsoft said the malicious version was never distributed through the Visual Studio Marketplace or npm. Cursor says it removed the extension, rebuilt its marketplace after rotating keys, and worked with the hosting provider to take the attacker's server offline.

ClaimStatus
A trojanized anthropic.claude-code 2.1.313 was served from Cursor's marketplace on October 29Confirmed by Cursor's notice
Up to 5,000 users may have installed itCursor's estimate, an upper bound
Anthropic's own systems were compromisedDenied by Anthropic
The build was on Open VSXEarly reporting; superseded. Anthropic says it was never published there, and Socket reported no evidence it was public there
The npm package @anthropic-ai/claude-code is affectedNo, per Anthropic and Microsoft
The C2 server is offlineStated by Cursor
How the attacker gained publishing access to Cursor's marketplaceNot disclosed

Why the Open VSX attribution spread

The first public reports on October 30 named Open VSX, the Eclipse Foundation registry that Cursor and other VS Code forks draw extensions from, and showed only a macOS build. Many write-ups, including the version that circulated on crypto news feeds, repeated that framing. Cursor's later notice placed the build in its own marketplace, and Anthropic and Microsoft both confirmed the official channels were clean. The confusion is understandable: from a developer's seat, an extension in Cursor looks like it came "from Open VSX", because Cursor pulls extensions from the Open VSX Registry for most listings. For incident response the distinction changes the scope. VS Code users who installed the extension from the Microsoft Marketplace are not in scope for this incident; Cursor users are.

What the implant does

The malicious build is the real extension with a loader of about 66 lines inserted ahead of the legitimate code, so the extension still works and nothing looks broken. When the extension activates, the loader:

  • Appends an attacker-controlled Ed25519 SSH public key with the comment svc-telemetry to the user's ~/.ssh/authorized_keys.
  • Tries to enable SSH remote login on the machine.
  • Opens a reverse SSH tunnel to 134.209.82.225, which reconnects if dropped, so the attacker can reach the laptop from outside even behind NAT.
  • Sends host reconnaissance to the same server. Phorion reports this includes the names, not the values, of up to 40 environment variables that look like keys, tokens, passwords, or AWS and GCP credentials: a map of what is worth stealing once the attacker is on the box.
  • Installs per-OS persistence that runs independently of the editor (see the table below).

The design choice to read only variable names is worth noticing. It keeps the first-stage traffic small and less alarming, while the SSH key gives the attacker interactive access to read the values, and everything else in the user's home directory, at a time of their choosing. Phorion also reported a bug in the implant's macOS setup logic, so not every install may have completed every step. Do not rely on that: check each machine.

Indicators of compromise

These indicators are drawn from the published analyses by Microsoft, Phorion, OpenSourceMalware and StepSecurity. Sources differ slightly on minor artifact names, so verify against Cursor's notice and Microsoft's write-up before relying on any single list.

WhereIndicator
Extensionanthropic.claude-code version 2.1.313 installed in Cursor (extension folder named anthropic.claude-code-2.1.313-* under ~/.cursor/extensions/)
SSHA key in ~/.ssh/authorized_keys with the comment svc-telemetry
NetworkConnections to 134.209.82.225, reported on ports 22 and 8099
macOSLaunchAgent ~/Library/LaunchAgents/com.apple.sysmond.plist (a user-level agent, not an Apple component); files ~/.config/.smt and ~/.config/.smtk; Remote Login switched on
WindowsScheduled task SysMon; %APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\sysmon.cmd; %USERPROFILE%\.config\smt.cmd and smtk; .stk_ files under %TEMP%
LinuxAn @reboot entry in the user's crontab that restarts the tunnel
Extension bundleThe strings 134.209.82.225 or svc-telemetry inside the installed extension's JavaScript

What to do if a machine ran 2.1.313

  1. Scope first. List every machine with Cursor installed and check the installed version of anthropic.claude-code. The window is October 29 from 09:27 to 20:35 UTC, but a machine that downloaded 2.1.313 in that window and has not updated since is still carrying it. VS Code installs from the Microsoft Marketplace are out of scope.
  2. Isolate and collect before cleaning. Take the machine off the network, then preserve authorized_keys, the persistence files, the crontab or task list, shell history and SSH logs. Removing the backdoor first destroys the record of whether the tunnel was ever used.
  3. Remove the key and the persistence, then the extension. Delete the svc-telemetry line from authorized_keys, remove the LaunchAgent, scheduled task, Startup script or cron entry, kill any running tunnel process, turn Remote Login back off if it was off before, and update the extension to 2.1.314 or later. Cursor has published removal scripts that do these steps and exit non-zero if anything remains; get them from Cursor's own notice, not from a link in an email or chat.
  4. Rotate from a clean machine. Assume anything readable by the user was readable by the attacker: SSH private keys, cloud CLI credentials, Git tokens, package-registry tokens, .env files, and the AI provider keys a coding agent's machine almost always holds. Rotate them from a machine that never ran the build.
  5. Block the indicator. Block 134.209.82.225 at the egress firewall and hunt for past connections to it in network logs, even though the server is reported offline.
  6. Check what the user's credentials reached. Review cloud audit logs, Git provider audit logs and SSH access logs on servers the user can reach for activity from October 29 onward.

Why this is an agentic-endpoint risk

An AI coding agent's extension is the highest-value impersonation target on a developer machine. It is expected to run shell commands, read the whole workspace, and talk to the network, so a loader hiding inside it inherits a reason for every suspicious thing it does. It is also updated often, which makes a new version number unremarkable. EDR sees a signed editor process loading its own extensions. The network sees an SSH connection from a developer laptop, which is ordinary. Nothing in the existing control stack is built to ask "which build of which AI extension is on this laptop, and where did it come from?"

The other lesson is that the marketplace is part of the trust chain, and VS Code forks each run their own. The same extension ID can resolve to different bytes in the Microsoft Marketplace, Open VSX and an editor vendor's own catalogue. A policy that approves anthropic.claude-code by name approves whatever any of those catalogues serves under that name. This is the same lesson as the Amazon Q Developer wiper prompt, where an official extension shipped attacker content, and it is one of the install-path patterns the AI supply-chain defender's guide tracks.

How Anomity helps

The first question in this incident is a fleet question: which machines have Cursor, and which of them carry anthropic.claude-code 2.1.313? Anomity's Endpoint Sensor reads the manifest of every installed Cursor extension and records its publisher, ID and version per device, alongside the AI agents, MCP servers, plugins, skills, hooks, CLIs and secrets on the same machine. That turns scoping from a survey of developers into a query, and the 90-day audit trail of version changes shows when each machine moved onto 2.1.313 and when it moved off, which is what the evidence window needs. Findings route to your SIEM, Slack, email or Jira.

Anomity also answers the follow-up question: once a machine is in scope, which secrets and AI provider keys were sitting on it. The sensor reports which credentials are present by type and location, with values redacted on the endpoint, so the rotation list in step four starts from a record rather than from memory.

It is worth being plain about the limits. This loader runs inside the editor's extension host, not as an agent tool call, so hook enforcement at Claude Code's PreToolUse event would not have intercepted it. Anomity does not read authorized_keys or replace EDR's view of the tunnel process. Anomity complements EDR, network and DLP tooling: they see the process and the connection, and Anomity tells you which AI extension builds are on which machines, which is the part they were never built to inventory.

You can't govern what you can't see.The Anomity principle

What to check across your fleet

  • Every machine with Cursor: the installed version of anthropic.claude-code. Anything on 2.1.313 is in scope; 2.1.312 and 2.1.314 are reported clean.
  • Every in-scope machine: grep svc-telemetry ~/.ssh/authorized_keys, and the per-OS persistence in the IOC table.
  • Egress and proxy logs: any connection to 134.209.82.225 from October 29 onward.
  • Which other VS Code forks are in use, and which marketplace each one pulls extensions from. The same extension ID is a different supply chain in each.
  • Whether extension auto-update is on in Cursor and other forks, and whether you want new versions of AI extensions held for a short delay.
  • The secrets and AI provider keys on in-scope machines, so rotation covers all of them.
  • For the wider coding-agent rollout, the inventory step in the fleet guide to governing AI coding assistants.

The trojanized 2.1.313 build is a small incident by install count and a large one by precedent: the official extension of a major AI coding agent, under its real ID, delivered by an editor's own update path, with a backdoor that outlives the extension. Check Cursor installs for 2.1.313, hunt for the svc-telemetry key, clean up and rotate. Then make the question "which AI extension builds are on which machines" one you can answer in minutes next time. To see Anomity inventory your AI extensions and agents, book a 30-minute demo.

Frequently asked questions

Was the Claude Code VS Code extension 2.1.313 on Open VSX backdoored?

Not according to Anthropic. Early reports on October 30, 2026, named Open VSX, but Anthropic says version 2.1.313 was never published to Open VSX, the Microsoft Visual Studio Marketplace, npm or any of its official hosts. Cursor's own notice places the trojanized build in Cursor's extension marketplace, where it was available for about 11 hours on October 29. Cursor pulls most extensions from Open VSX, which is why the two were easy to confuse. If you installed Claude Code in Cursor, you may be affected; if you installed it in VS Code from the Microsoft Marketplace, you are not affected by this incident.

How do I know if my machine has the Claude Code extension backdoor?

Check three things. First, whether Cursor has anthropic.claude-code version 2.1.313 installed, or did at any point since October 29, 2026. Second, whether ~/.ssh/authorized_keys contains a key with the comment svc-telemetry. Third, whether the per-OS persistence exists: a com.apple.sysmond.plist LaunchAgent on macOS, a SysMon scheduled task or sysmon.cmd Startup script on Windows, or an @reboot cron entry on Linux. Any of these means the machine should be treated as compromised.

Does uninstalling the Claude Code extension remove the backdoor?

No. The loader installs an SSH key and standalone persistence that run independently of the editor, so removing or updating the extension leaves the backdoor in place. Collect evidence first, then remove the svc-telemetry key, the LaunchAgent, scheduled task or cron entry, and any running tunnel process, and rotate every credential the user could read, from a clean machine. Cursor has published removal scripts in its own notice.

Which Claude Code extension versions are safe?

Version 2.1.313 from Cursor's marketplace is the malicious one. Anthropic skipped to 2.1.314 to avoid confusion, and 2.1.312 and 2.1.314 are reported clean. The npm package @anthropic-ai/claude-code and the builds on the Microsoft Visual Studio Marketplace and Open VSX were not affected, per Anthropic and Microsoft.

How does Anomity help with this?

Anomity's Endpoint Sensor records every installed Cursor and VS Code extension with its publisher, ID and version on each managed machine, so finding every device that carries anthropic.claude-code 2.1.313 is a query rather than a survey. Its 90-day audit trail shows when each machine moved onto and off the bad build, and its secrets inventory, with values redacted on the endpoint, shows which credentials were on in-scope machines and need rotating. Anomity complements EDR and network tooling, which see the backdoor's process and connections.

Ask AI about Anomity
ChatGPT Claude Perplexity Google AI Grok