Skip to content

Our article2 min read

Cursor Says a Fireworks Key Led to the Fake Claude Code Extension: What to Check

Cursor says a key exposed by Fireworks AI let attackers slip tampered Claude Code builds into its marketplace on 29 October. Here is how the attack is described, the figures, and how to check your machine.

Maya, the All About AI writer

By Maya, AI writer
Checked by a person at Zettaz before publishing

ShareFacebookXLinkedIn
In this article

Key takeaways

  • Cursor says tampered Claude Code versions 2.1.312 and 2.1.313 were served for about 11 hours on 29 October, reaching at most 5,000 users.
  • In a 13 November update, as reported, Cursor traced the attack to an access key it had issued to Fireworks AI, which Fireworks later exposed.
  • Uninstalling the extension may not be enough: it tried to add an SSH key labelled svc-telemetry and set up persistence.
  • Anthropic says it did not build or publish the tampered copy.

What happened on 29 October

Cursor's security notice says versions 2.1.312 and 2.1.313 of the Claude Code extension, for Windows and macOS, were served from its marketplace for roughly 11 hours on 29 October 2026, from about 2:27am to 1:35pm Pacific time. Cursor says the malicious versions were not developed by Anthropic. Anthropic told reporters, as cited by BankInfoSecurity, that it did not build, publish or host the tampered extension and released a clean version, 2.1.314, that evening.

What the code tried to do

Per Cursor's notice, the tampered extension collected host details such as the username, operating system and public IP address, and tried to add an attacker-controlled SSH key, switch on the local SSH service and open a reverse tunnel. It also tried to keep itself running through a macOS LaunchAgent or a Windows scheduled task. Cursor says the attacker's server is offline and that the SSH steps would usually fail without administrator rights, but the key, persistence and host data collection could still happen, so affected machines need cleaning.

The numbers, and what they do not show

  • At most 5,000 users could have received the tampered build, a potential count rather than confirmed infections.
  • Fewer than 0.05% of users had traces exposed, and fewer than 0.01% of customers had API credentials exposed.
  • Cursor gives percentages, not absolute customer numbers, and says Privacy Mode data and its official updates were unaffected.
  • Fireworks has not published how its key was exposed, per Data Studios, and its own investigation is still open.

How to check your machine

If you had Claude Code 2.1.312 or 2.1.313 in Cursor on 29 October, Cursor says to follow its remediation steps. On a Mac you can run grep -qs svc-telemetry ~/.ssh/authorized_keys; an exit status of 0 means the attacker's key is present. On Windows, search both the user and administrators authorized_keys files for the same text. If found, treat the machine as compromised: isolate it, remove the key and persistence files, rotate credentials reachable from it using a trusted computer, and consider rebuilding. Cursor says users who received no notice were not in the affected traces.

What it means for you

Most people never installed the affected versions. The wider lesson is that editor extensions run with your own permissions, and keys passed to a vendor can end up in data you never see. Keep a list of which services hold copies of your keys so one notice lets you rotate quickly. Our guide to using AI at work safely covers related habits. This summary relies on Cursor's statements as reported, not on our own testing.

Go deeper

Sources

Written by an AI from the sources above and checked by a person before publishing. We have not tested products unless we say so. See our editorial policy or tell us about an error through contact.

ShareFacebookXLinkedIn

Keep reading