AI incidents rarely look like a classic breach. A customer-facing assistant gives harmful advice, a model upgrade silently degrades a credit decision, an agent follows instructions planted in a document and emails data out, or a meeting summariser leaks a board discussion. Each needs a responder who can tell whether it is a security, safety, privacy, rights or quality event, contain it without destroying evidence, and work out which reporting clocks started when.
You will define AI incidents, hazards and near misses using the OECD definitions, build a taxonomy and an AI annex to an incident response plan aligned with NIST SP 800-61r3, and specify the logging needed before anything goes wrong: prompts, outputs, tool calls, retrieval sources and model versions. You will write playbooks, use guardrail alerts, evaluation regressions and MITRE ATLAS techniques as detection signals, classify severity, and contain incidents through feature flags, model rollback and credential or tool revocation. Investigation modules cover root cause analysis when outputs are not reproducible, evidence preservation and privilege, and re-release criteria. You will map overlapping notification duties, including EU AI Act Article 73 serious incident reporting, GDPR breach notification, NIS2, DORA, SEC Form 8-K Item 1.05, and frontier-developer incident reporting under California SB 53 and New York's RAISE Act. The course ends with blameless reviews, disclosure and tabletop exercises. Scenarios are hypothetical; educational, not legal advice. The Wirewalk Institute certificate attests to completion and assessment of this course only.