Cursor Traces Its Marketplace Malware to a Vendor Key That Exposed Old Coding Agent Sessions
Cursor says a key exposed by Fireworks opened access to historical agent traces. Credentials in those records helped an attacker reach its marketplace cache.
Cursor has explained how an attacker reached the marketplace cache that served developers a malicious Claude Code extension: a vendor key opened the door to old AI coding sessions, and credentials left in those sessions opened more doors.
In its November 13 investigation update, Cursor says Fireworks exposed an access key it had issued to the vendor. An attacker used it to copy historical agent traces, mostly from 2025, from a shared storage bucket. Some contained unredacted credentials.
Those credentials led into internal software services. A stolen credential then let the attacker put malicious extension builds directly into Cursor's marketplace cache on October 29. Cursor says Open VSX never hosted the malware and Anthropic did not publish it.
Old conversations became a route inside
Cursor puts exposed API credentials at less than 0.01% of customers. That figure describes the company's investigation, not an independent audit.
The newly described entry point answers a question left open in our coverage of the original marketplace attack: how the attacker got access to distribution.
Removing an extension can leave the payload behind
Cursor's original security notice identifies the affected builds as Anthropic.claude-code 2.1.312 and 2.1.313. The payload attempted to install an SSH backdoor and persistence on Windows and macOS. Coding errors and missing administrator privileges limited some steps, but the notice says an attacker's key, persistence files and host information could still be left behind.
That is why Cursor directs affected users to its detection and remediation instructions. Uninstalling the extension alone may leave separate scripts running.
The chain turns records kept for model development into a security problem for the software used to build those models. A secret in an old conversation can still open a current system.
Some offers on this page may be paid placements or contain affiliate links.
Cyberpresso: daily cyber & AI brief
Free daily newsletter, read in 5 minutes.
Subscribe free