The first trust-chain correction matters: Cursor says the attacker obtained a Cursor-issued access key entrusted to Fireworks, not a key issued by Fireworks to Cursor. Verified: that key accessed a read-only bucket containing historical coding-agent traces; some traces exposed usable credentials; a credential then enabled malicious publication through Cursor’s extension marketplace. Cursor identifies malicious Anthropic.claude-code 2.1.313, while Microsoft reports both 2.1.312 and 2.1.313 and confirms an SSH-persistence payload. Cursor’s documented serving window—2:27 a.m. to 1:35 p.m. PT on October 29—is approximately 11 hours. Microsoft also confirms that Anthropic’s npm distribution and Microsoft’s Visual Studio Marketplace were unaffected. Upper bound: “up to 5,000 users” describes potential distribution exposure, not verified installation, execution, persistence, or interactive attacker access. Those host-level counts remain unknown.
The observed trust failures were sequential: third-party custody of Cursor’s key; a shared data bucket containing traces with unredacted secrets; secondary credentials retaining usable authority; inadequate separation between compromised internal access and marketplace publishing; publisher authentication without independent artifact provenance; marketplace ingestion accepting a trojanized Anthropic-branded build; cache and revocation controls permitting continued distribution during the exposure window; and developer workstations trusting marketplace identity sufficiently to execute code in a credential-rich environment. This was Cursor’s distribution path—not Anthropic’s upstream release channel—and authenticated publication was mistaken for trustworthy provenance.
For each potentially affected workstation, isolate it and preserve memory, disk, extension directories, IDE logs, network telemetry, and installation timestamps. Establish whether 2.1.312 or 2.1.313 was merely downloaded, installed, or loaded; inspect authorized_keys for the reported svc-telemetry key, SSH configuration and services, persistence scripts, reverse-tunnel activity, and subsequent commands or files. Maya Chen’s minimum trust bar is clear: uninstalling the extension is insufficient; if it executed or telemetry is incomplete, reimage from known-good media. From a clean device, revoke sessions and rotate developer SSH keys, source-control and package-registry tokens, cloud/API credentials, CI/CD secrets, signing keys, and any credentials accessible to the IDE or SSH agent. Cursor must revoke publisher credentials, audit every publisher and version change, purge CDN caches, block the malicious hashes, identify recipients, and require artifacts cryptographically tied to the legitimate upstream publisher before relisting.
Compared with N-central, this has stronger evidence of a malicious artifact and host persistence but a narrower demonstrated distribution scope. N-central carries the larger systemic MSP blast radius and therefore justifies platform isolation even without observed exploitation. Cursor demands greater endpoint urgency for hosts that installed or may have executed the malicious builds: rebuild and rotate secrets host by host. The plans should remain separate—N-central is a privileged management-control-plane containment problem; Cursor is a developer endpoint, credential, publisher, and marketplace-provenance incident.