AI Incident Response: LLM & Agent Failures in Production

This course contains the use of artificial intelligence.It is four in the morning. Your autonomous agent has just sent fourteen hundred customers a letter nobody approved. Uptime is fine. Latency is fine. The error rate has been zero all night. Every dashboard you own is green — and that is exactly the problem. This is an operations course. Not governance, not red teaming, not architecture. The alarm has already fired, and you are the one holding the pager. You will learn to detect, triage, contain, preserve evidence, investigate and recover from failures in deployed LLM and agent systems — organised by failure mode, never by framework. Everything is taught against Meridian Health, a fictional insurer running an LLM claims assistant and an autonomous operations agent with tool access to a claims database, outbound email and an internal MCP server. Every incident you respond to happens to Meridian first. What makes this course differentBuilt on the Coalition for Secure AI (CoSAI) AI Incident Response Framework V1.0 — the first genuinely authoritative reference in this field — and made operational rather than summarised. Eleven named playbooks, each following the same five-part spine: signals, containment, evidence, eradication, recovery gate. Real cases, including the Canadian tribunal decision that rejected "the chatbot is a separate legal entity" defence, and the documented 2026 runaway-cost incidents. Cost-as-an-incident is covered properly — under-developed even in the frameworks, and the best-documented failure mode of 2026. You will leave with five artefacts you can put into production: an AI incident severity matrix, containment procedures for your own stack, your own incident playbook, an evidence checklist with chain of custody, and a model-aware post-incident review template. The capstone assembles all five with a gap statement and a 90-day roadmap. Nine hands-on labs run entirely on your own machine against a local model — no API keys, no cloud spend. You will instrument an agent, score live incidents, run a timed containment drill, reconstruct an incident from raw evidence, and rebuild an eval gate that refuses to open until the fix is proven. No attack-development experience is needed. Every incident in the labs is handed to you already in progress. You are the responder, never the attacker.