[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

official (Cursor incident update, Nov 13, 2026); Cursor says contained, last attacker activity Nov 1

Cursor traces its October breach to a key it gave Fireworks AI: stolen coding-agent traces held unredacted credentials, which led to SaaS tools, support data and the fake Claude Code extension

  • The chain: vendor-held key → training-data bucket → secrets pasted into agent sessions → SaaS tools → a credential that could write to the extension cache. No zero-day was involved, per Cursor.
  • Cursor's fixes: moved all inference off Fireworks and cut its data access, rotated every exposed credential and, as a precaution, all internal ones, rebuilt the extension marketplace, restricted where service-provider credentials can be used, added secret scanning to repos and other internal data, redeployed production, and brought in external forensics.
  • If you used Cursor on Oct 29: follow the Oct 29 notice for Anthropic.claude-code 2.1.312/2.1.313 (svc-telemetry key check); uninstalling alone may not be enough.
  • If you were notified: rotate the keys listed in Cursor's email; any secret you pasted into an agent session or sent to support should be treated as exposed.
  • Still unknown: whether any exposed customer key was misused, and absolute customer counts.

Sources: Cursor: An update on our recent security incident (Nov 13, 2026) · Cursor: Security notice, malicious Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29-30, 2026) · BleepingComputer: Cursor confirms security breach led to malicious Claude Code extension (Nov 14, 2026) · Fireworks AI: Update on October 2026 security incident (Nov 9, 2026)