Cursor, the company behind the AI code editor of the same name, published a follow-up report dated November 13 on the malicious "Claude Code" extension that was distributed through its VS Code extension marketplace on October 29. According to the company, the intrusion began with an access key exposed on the side of Fireworks, the provider it used for inference. From there, the attacker got into Cursor's internal tools and used a stolen credential to swap in the tampered extension.

The Malicious Extension Was Live for About 11 Hours and May Have Reached Up to 5,000 Users

The affected builds posed as versions 2.1.312 and 2.1.313 of Anthropic's Claude Code extension (Anthropic.claude-code) for Windows and macOS. Cursor's notice says they were available for about 11 hours, from around 6:27 p.m. on October 29 to around 5:35 a.m. on October 30 Japan time (JST), or roughly 2:27 a.m. to 1:35 p.m. Pacific Time on October 29. Up to 5,000 users who requested those versions during that window may have received the malicious build.

Cursor says Anthropic did not publish this copy and that the incident was not caused by a vulnerability in Anthropic's product. Open VSX, the source for the marketplace, never contained the malicious code either; the attacker placed the copies directly on the cache of Cursor's marketplace.

The Malicious Code Aimed for an SSH Backdoor and Persistence

According to the notice, once run, the malicious extension collects information about the host and user and sends it to the attacker. It also tries to add an SSH public key labeled "svc-telemetry" to authorized_keys, enable SSH, and open a reverse tunnel. On top of that, it registers itself as a LaunchAgent on macOS or as a scheduled task on Windows and tries to run again roughly every 60 seconds.

Because enabling SSH requires administrator privileges, Cursor expects that step to fail on most machines. The company says the attacker's server no longer functions.

It Started With an Access Key Exposed at the Inference Provider

According to the follow-up report, the attacker used an access key Cursor had issued to Fireworks to get into storage the two companies used to share data. That storage held limited coding agent traces, mostly from 2025, from users who were not using Privacy Mode. Some traces contained unredacted credentials, which the attacker used as a foothold to break into SaaS tools Cursor uses for its business.

The information Cursor lists as exposed includes archived non-production source code repositories (some containing business contact information), some support records and email conversations, product usage metrics and contact details, summaries of some agent conversations, and credentials users had sent to support. Cursor says traces were exposed for fewer than 0.05% of users, and API credentials for fewer than 0.01% of customers. The last attacker activity it has seen was on November 1, and it was blocked.

Cursor says its production systems, the Cursor app and official updates, release signing keys, customer code repositories, the prompts and code of Privacy Mode users, and other extensions were not affected.

Cursor's Response and What Users Should Check

Cursor moved all inference traffic off Fireworks and deactivated the key, and rotated not only the exposed credentials but also all of its internal credentials as a precaution. It reviewed and rebuilt the extension marketplace, expanded network access restrictions, added secret scanning across its code repositories, and redeployed its production infrastructure. It has also brought in outside forensics experts and contacted affected customers individually.

Anyone who may have installed or updated to the affected versions during the exposure window should follow the steps in the notice and check whether a "svc-telemetry" key is present in authorized_keys. If it is, Cursor says to treat the machine as compromised: isolate it from the network, remove the key, the persistence settings, and the tunnel processes, and consider rebuilding the machine. Cursor notes that uninstalling the extension alone may not be enough. It advises rotating credentials and sessions that were accessible from that machine using a separate, trusted system, and provides remediation scripts for macOS and Windows on the notice page.

Our Take

This incident shows that the more AI development tools rely on outside inference services and data sharing for training and analysis, the more a single key at a vendor can become a path all the way to users' machines. The fact that credentials were left intact in agent traces is not someone else's problem for users of other AI coding tools, either. Beyond trusting where extensions come from, teams will likely need practices that limit how many secrets they hand to agents.

Summary

Cursor has disclosed that the malicious Claude Code extension distributed on October 29 resulted from an intrusion that began with an access key exposed at its inference provider, Fireworks. The malicious builds were available for about 11 hours and may have reached up to 5,000 users. Anyone who may have installed the affected versions should check their machine using the steps in Cursor's notice and rotate their credentials.