Skip to main content

Incident response playbook

Stub

This How-to is a stub. A full incident response playbook is planned, starting with prompt injection, data leaks, and runaway loops.

Goal​

You will respond to common incidents: prompt injection, data egress, and runaway execution.

When to use this​

  • You suspect policy bypass or data leakage.
  • A run is looping or causing unexpected cost.

Prerequisites​

  • Durable artifacts/events enabled
  • Ability to revoke credentials/secrets quickly
  • A rollback strategy (JARVIS_Release pinning)

Steps​

  1. Contain:
    • stop the affected runs,
    • revoke tokens/secrets if needed,
    • tighten policy temporarily (deny more).
  2. Triage using artifacts:
    • candidate sets,
    • binding decisions,
    • tool calls / outputs.
  3. Remediate:
    • patch the capability,
    • add evaluators/policy gates,
    • roll forward with a pinned release.

Verify​

  • The incident scope is understood and contained.
  • A postmortem exists with actionable follow-ups.

Troubleshooting​

  • Missing evidence → increase artifact coverage before returning to prod.
  • Unclear root cause → add tracing and better decision artifacts.
  • Repeated incidents → tighten action space (candidate sets, budgets) and add policy gates.

Cleanup / Rollback​

  • Rotate secrets and roll back to the last known-good release bundle.

Next steps​