From inference provider to IDE backdoor: the Fireworks and Cursor attack chain
From Inference Vendor to IDE Backdoor: The Fireworks–Cursor Chain
The Cloud Security Alliance reviewed the October 29 incident in which the Cursor extension marketplace distributed a tampered Anthropic Claude Code extension, noting that the chain began with Cursor's inference and training provider Fireworks AI. According to Cursor's November 13 statement, attackers obtained an access key Cursor had issued to Fireworks, copied a read-only storage bucket shared with Fireworks—most of the coding agent session records in the bucket were from 2025 and some still contained unredacted credentials—and these credentials entered SaaS tools used by Cursor, one of which was used to write a malicious build into the marketplace cache. The affected versions are Anthropic.claude-code 2.1.312 and 2.1.313, and Cursor estimates that up to 5000 users requested the affected cached versions during that window.
CSA breaks the backdooring of the Cursor extension into four trust boundaries, explaining how a key leaked by an inference provider travels all the way to the developer endpoint.