The Agentic Post
Breaking
Black October 2026: Every Major AI Security Incident, Explained  ·  ARTEX: The Open-Source AI Hacking Tool Behind October’s Biggest Attacks  ·  Chinese Hackers Used AI Agents to Hit 2,250 Microsoft 365 Accounts  ·  PHP2Shell Is Now Being Exploited, and Attacks Began Before the Patch  ·  OpenAI-Backed Thrive Buys ADT for $15.3 Billion to Rebuild It With AI  ·  The AI Data Center Backlash Hits Minnesota and Australia in One Week  ·  
Home/AI Safety
Fake Claude Code Extension Planted SSH Backdoors on Up to 5,000 Machines

Fake Claude Code Extension Planted SSH Backdoors on Up to 5,000 Machines

AI Safety

A trojanized Claude Code extension, version 2.1.313, sat in the Open VSX store used by Cursor for 11 hours, planting SSH keys and a reverse tunnel. Up to 5,000 users affected. Update to 2.1.314.

A trojanized copy of the official Claude Code extension, version 2.1.313, was available in the Open VSX marketplace used by Cursor and other VS Code-based editors for about eleven hours on October 29, 2026. Simply loading it planted an attacker’s SSH key on the machine, opened a hidden remote tunnel and set itself up to survive reboots. Cursor says at most 5,000 users were affected. Anyone who installed 2.1.313 should treat their computer as compromised. A clean version, 2.1.314, was published on October 30.

What did the fake extension do?

It looked and worked like the real Claude Code extension, but carried a small hidden loader. In plain terms, it gave a stranger a permanent back door into your computer:

  • Planted a key: added an attacker-controlled SSH key, labelled svc-telemetry, to your ~/.ssh/authorized_keys file, which lets someone log in to your machine remotely.
  • Opened a tunnel: started a reverse SSH connection to an outside server so the attacker could reach your machine even behind a home router.
  • Stayed hidden: installed persistence disguised as a system service: a scheduled task called "SysMon" on Windows and a LaunchAgent named com.apple.sysmond on macOS.

No click or approval was needed beyond the editor loading the extension. The attacker’s server has since been taken offline by its hosting provider, but any key or task already planted stays until you remove it.

Am I affected?

Only if you installed version 2.1.313 from Open VSX, the extension store used by Cursor, Windsurf, VSCodium and similar editors, during the window on October 29. Users of the official Microsoft VS Code Marketplace were not the reported target. To check, open your extensions list in Cursor or run cursor --list-extensions --show-versions and look for anthropic.claude-code@2.1.313.

What should I do if I installed it?

  • Update the extension to 2.1.314 or later.
  • Open ~/.ssh/authorized_keys and delete any line ending in svc-telemetry or any key you do not recognise.
  • Remove the persistence: on Mac, delete ~/Library/LaunchAgents/com.apple.sysmond.plist. On Windows, delete the "SysMon" scheduled task in Task Scheduler.
  • Turn off remote login if you never use it (on Mac: System Settings, General, Sharing, Remote Login).
  • Rotate secrets on that machine: SSH keys, GitHub tokens, cloud keys and AI API keys, including your Anthropic key.
  • If in doubt, reinstall the operating system. Cursor’s own advice is to treat the machine as compromised.

Why do attackers keep targeting AI coding tools?

Because a developer laptop running an AI agent is a master key. It holds source code, cloud credentials and API keys, and the agent already has permission to run commands. This is the latest in a run of supply-chain attacks aimed at exactly that setup, including the poisoned LiteLLM package and GitSpawn. Open VSX has been hit repeatedly this year because it has lighter review than Microsoft’s marketplace and is the default store for several popular AI editors.

The lasting lesson: an extension carrying a trusted name is not proof it came from that company. Turn off automatic extension updates for tools that hold your keys, and check release notes before updating.

Part of our coverage of Black October 2026, every major AI security incident this month.

Up Next
How to Read AI Comparisons

How to Read AI Comparisons

Comparisons

A practical guide to reading AI product comparisons critically, covering what's actually being measured, publish date, and testing finalists yourself.

Comparison articles, including the ones on this site, are genuinely useful starting points and shouldn’t be the final word on which AI tool to use. Here’s how to read them well.

Check what’s actually being compared. A comparison based on general benchmark scores measures something different from one based on hands-on testing for a specific task. Neither is wrong, but knowing which kind you’re reading changes how much weight to give its conclusions.

Watch the publish date closely. Our full guide to reading benchmark scores covers how fast this field moves: a comparison from even a few months ago may already be outdated given how frequently leading models release new versions.

Use comparisons to narrow options, not to make the final decision. A good one should get you from a dozen options down to two or three worth actually testing yourself, not hand you a single definitive winner.

Once you’re down to real candidates, run your actual recurring task through both for a real week before committing. This consistently reveals more than any article can. See Terminal-Bench for independent data as one input.