Cursor confirmed on November 13 that a stolen Fireworks AI access key set off a chain of breaches that ended with a backdoored version of Anthropic’s Claude Code extension sitting in its VS Code marketplace for about 11 hours on October 29. Up to 5,000 users who installed or updated the extension on Windows or macOS during that window may have received the tampered build, which tried to plant a persistent SSH backdoor on their machines.
How a Third-Party AI Key Turned Into a Backdoored Extension
According to Cursor’s own security incident update, the trouble started outside Cursor entirely. Fireworks AI, a third-party model provider Cursor used for inference, exposed an access key it had issued to Cursor. An attacker used that key to copy data out of a read-only storage bucket Cursor shared with Fireworks, mostly coding-agent traces from 2025 that belonged to users who hadn’t turned on Privacy Mode.
Some of those traces still had unredacted credentials sitting in them. The attacker used those leaked credentials to reach a handful of Cursor’s internal SaaS tools, and from there got a credential that let them push malicious copies of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313, into the cache serving Cursor’s VS Code extension marketplace. Cursor is explicit that Anthropic never published these builds; the marketplace itself was the weak link.
What the Tampered Extension Actually Did
Per Cursor’s security notice, the fake builds were live between roughly 2:27am and 1:35pm Pacific time on October 29. Once installed, the code collected host details, including hostname, OS, architecture, working directory, Git email, and public IP, and tried to drop an attacker-controlled SSH key labeled svc-telemetry into authorized_keys. It then attempted to enable SSH and open a reverse tunnel, with persistence set up through a macOS LaunchAgent or a Windows scheduled task. Cursor notes this step generally needs admin rights to succeed, which likely limited how many machines actually ended up with a working backdoor.
Who’s Actually in the Blast Radius
The numbers are narrower than the headline suggests. Cursor says fewer than 0.05% of non-Privacy-Mode users had coding-agent traces exposed, and fewer than 0.01% of customers had API credentials compromised. The attacker’s last observed activity was November 1, and Cursor says it was blocked. Production infrastructure, official app releases, release signing keys, other marketplace extensions, and anything touched by Privacy Mode users were not affected, per the company.
This is the second time in a few weeks that AI coding agents have shown up as an attack surface rather than just a productivity tool. We wrote about a similar pattern last month when malicious AGENTS.md files started tricking Codex and Claude Code into installing malware, and before that in the Plugin4Shell flaw that hit Claude Code, Codex, Copilot, and Gemini CLI. The common thread is the same: the attack rarely targets the model itself, it targets the plumbing around it, extension marketplaces, third-party inference providers, config files an agent trusts by default.
What to Do if Your Team Runs Cursor
If anyone on your team installed or updated the Claude Code extension in Cursor on October 29, Cursor’s notice is blunt about one thing: uninstalling the extension alone may not be enough if the backdoor was actually planted. Its recommended steps:
- Check for the
svc-telemetrySSH key inauthorized_keyson any Windows or macOS machine that had the extension during the exposure window. - Run Cursor’s platform-specific remediation script (
remediate-macos.shorremediate-windows.ps1) rather than assuming a clean reinstall fixed it. - Rotate credentials and sessions that were reachable from an affected machine, starting from a trusted, uncompromised device.
- Treat any host where the payload actually ran as needing a rebuild, not just a cleanup.
For teams building AI tooling into client products rather than just using it personally, this is also a reminder to extend normal vendor-risk thinking to AI coding assistants and their extension marketplaces, not just to the SaaS platforms you already audit. ConicPlex’s AI Development work includes exactly this kind of review when AI tools get woven into a client’s actual development pipeline.

