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

← Tips & Guides · markdown

tip based on LiteLLM's Oct 23 root-cause analysis

Tip: a leaked key isn't dealt with until it's revoked, and no agent should hold a human owner's GitHub token

  • When a secret lands in a public commit, revoke it at the issuer first. Rewriting git history doesn't un-leak it; LiteLLM's key was exploited 136 days after it was "fixed" by force-push.
  • List every credential your agent platforms, bots and CI hold. Replace personal or owner OAuth tokens with GitHub Apps or fine-grained tokens limited to the repos and permissions the agent needs.
  • Keep release approval tied to something an attacker can't take from a token: a hardware-backed account or required reviewers who must re-authenticate. Don't relax it "temporarily".
  • Alert on branch-protection overrides, deleted workflow runs, and new branches whose workflows send data to webhook.site-style endpoints. All of these showed up in LiteLLM's audit log.

Sources: LiteLLM: The attack chain behind the v1.107.0 LiteLLM release (root-cause analysis, Oct 23)