What happened
Confirmed supply-chain compromise of a widely-used AI coding agent's official extension. Fireworks AI was hacked twice in October; a Cursor key in its custody leaked, granting read access to a bucket of old coding-agent traces that still held unredacted credentials, then access to Cursor SaaS tools and finally the marketplace cache. Malicious Claude Code builds were served for roughly 11 hours; Anthropic did not build or publish them, and Open VSX was never affected. (Disclosed 2026-11-15/16.)
Why it matters
Shows how a third-party AI inference vendor's key and data retention become a software-supply-chain entry point into developer workstations — an official AI coding-agent extension distributing a persistent SSH backdoor, underscoring that AI-vendor risk includes the data exchanged for training/analysis.
Attack vector
An access key Cursor had issued to Fireworks AI was exposed in Fireworks' October breach; the attacker used it to copy unredacted credentials from a shared read-only storage bucket (retained coding-agent traces), pivot into Cursor's SaaS tools, and then place malicious copies of Anthropic's Claude Code extension on the cache serving Cursor's marketplace. The tampered extension profiled the host, planted an SSH key, attempted a reverse tunnel, and installed persistence (LaunchAgent / scheduled task).
Affected systems
anthropic.claude-code Cursor VS Code extension versions 2.1.312 and 2.1.313 (macOS/Windows), served via the Cursor extension-marketplace cache
Mitigation
Cursor's security notice: affected users must not simply uninstall — rotate exposed credentials, inspect authorized_keys and persistence items, and reinstall the clean 2.1.314. Cursor revoked the key, moved inference off Fireworks, and took the marketplace offline during the window. Note the persistence scripts survive an uninstall.