Skip to main content

AI-introduced defect incident runbook

Extends the standard incident flow with the RAPID fork: when a defect is found, ask "was this AI-generated?" — and if so, fix the class, not just the instance.

The fork​

  1. Standard incident response — contain, mitigate, restore (per SRE/ITIL).
  2. "Was this AI-generated?" — check the provenance trail.
  3. If yes — blast-radius query: what else did this agent / prompt / skill build? The provenance trail turns one defect into a query over everything that shares its origin.
  4. Patch the source: fix the offending prompt, skill, guardrail or standard so the defect class cannot recur — not just this instance.
  5. Deposit the capital: regression test, gotcha, and a blameless post-incident summary (templates).

Why this is different from a normal incident​

A human-written bug is usually one bug. An AI-generated bug may be a pattern — the same flawed prompt or skill could have produced the same mistake in a dozen places. The blast-radius query and the fix-the-source step are what stop a systematic defect from being patched one-by-one forever.

Standards referenced: NIST SP 800-61, blameless postmortems.