A trapped Claude Code extension targeted developers via Cursor

A modified version of the Claude Code extension was distributed for about 11 hours on the Cursor marketplace. It added an SSH key, persistence mechanisms, and a reverse tunnel.

A modified version of Anthropic.claude-code, the Anthropic Claude Code extension distributed via the Cursor marketplace, was online on October 29, 2026, for about 11 hours. Cursor says up to 5,000 users may have been exposed, with code capable of adding an SSH key, installing persistence, and opening a reverse tunnel. The risk is twofold: the attack came through a trusted extension and could leave traces after removal.

A marketplace for extensions that also becomes an entry point

Version 2.1.313 was not published by Anthropic but through Cursor’s distribution channel, the developer of a VS Code-based development environment. According to Cursor’s note, the malicious package was visible between 2:27 a.m. and 1:35 p.m. Pacific Time on October 29, 2026. The compromise does not affect the official edition of Claude Code, but rather the copy distributed by Cursor’s marketplace. This is a technical nuance, and it changes everything.

The malicious code began by collecting information about the host: machine name, username, operating system, architecture, working directories, public IP address, and other configuration details. Then, it added an SSH public key named svc-telemetry to the authorized_keys file. In other words, the machine could remain accessible to the attacker even after the extension was uninstalled.

Persistence, reverse tunnel, and incomplete cleanup

The second part was more discreet but longer-lasting. The package attempted to activate the local SSH service, then establish a reverse tunnel to 134.209.82.225 on port 22, using tools like ssh -N -R and autossh -M 0. Cursor indicates, however, that activating the SSH service required administrator privileges and failed in most cases. This does not negate the impact: the SSH key was indeed added, and persistence scripts were deployed.

On macOS, these scripts took the form of a LaunchAgent named com.apple.sysmond; on Windows, a scheduled task SysMon with a fallback in the startup folder. The mechanisms were designed to retry the connection approximately every 60 seconds. An extension can therefore do much more than load code; it can also embed itself in the environment, like a very polite squatter.

Cursor cut the server, but not necessarily the traces

Cursor claims to have removed the malicious version, rebuilt its marketplace, rotated its keys, and worked with the host to take the attacking server offline. The company also says it found no evidence of alteration of other extensions. At this stage, the external tunnel and beacon mechanism are no longer functional, but this is not enough to consider the machines clean.

The instruction is clear for users who installed or updated Anthropic.claude-code 2.1.313 during the affected window: check for the presence of the svc-telemetry key, search for persistence scripts and tasks, then isolate affected hosts. Cursor also recommends revoking and rotating credentials accessible from these machines, as the initial data collection does not close the door to later access via the backdoor.

What the incident reveals about risk in development tools

The case reminds us that a code editor, especially when it includes AI assistants and third-party extensions, concentrates a lot of trust in very few places. A single compromised update can modify SSH keys, create background tasks, and affect remote access to a development machine. For teams, the issue goes beyond simple software hygiene: it also requires monitoring the distribution chain, permissions granted to extensions, and indicators of compromise on workstations.

For now, the episode seems limited to a short window and a specific package, but it shows one very concrete thing: when an extension is compromised, the “uninstall” button is not always a magic button.

By the numbers

  • 11 hours — approximate exposure duration of the package on October 29, 2026, according to Cursor.
  • 5,000 users — maximum number of potentially affected users, according to Cursor.
  • 2:27 a.m. to 1:35 p.m. PT — estimated distribution window of package 2.1.313.
  • 134.209.82.225 — address of the attacking server identified by Cursor.
  • October 29, 2026 — date of discovery and alert publication.

To read