⭐ Star on GitHub

Summary: Wiki pages for raw/AI-incident-response/ — how incident response changes when the system under attack is probabilistic, agentic, and produces harmful output at machine speed.

Sources: raw/AI-incident-response/ (4 files, ~4.7k words)

Last updated: 2026-07-28


Pages

PageSource
ai-incident-responseSynthesis across all four sources — the concept, what breaks, containment, clocks
microsoft-ir-readinessMicrosoft Zero Trust / SFI pattern — three-stage remediation, trade-offs, success factors
cosai-frameworkCoSAI AI Incident Response Framework v1.0 — CACAO playbooks, ATLAS mapping, autonomy ladder
aiid-incident-responsesAI Incident Database — public disclosure responses and what they consistently omit

The term means three different things

  1. Using AI to accelerate IR (Wiz) — behavioural detection, automated timeline generation, AI-guided triage.
  2. Responding to incidents targeting AI systems (Microsoft, CoSAI) — the one this vault cares about.
  3. A public official statement by the responsible entity (AIID) — a disclosure artefact, not a process.

Meanings 2 and 3 share a name and nothing else. Keep them apart.

The core problem, in one sentence

“Traditional incident response assumes deterministic systems… A gap in a safety classifier can produce thousands of harmful outputs before any reviewer sees the first one” (source: raw/AI-incident-response/Incident response for AI systems.md).

Six things to take away

  1. Severity cannot be record count. Weight by “deployment domain, affected population, and content nature.”
  2. Root cause may not exist as a single defect. Behaviour emerges from training data × fine-tuning × retrieval × user context. Manage stakeholder expectations shaped by traditional defect analysis.
  3. A single test pass does not verify a fix. Watch periods after each remediation stage are mandatory, because the system is non-deterministic.
  4. Capture prompt logs, inference activity, tool executions, and memory state changes. CoSAI’s telemetry list. The last two are what agentic incidents are made of, and what most teams don’t log.
  5. Containment is architecture-specific — roll back the model, purge memory, or rebuild the vector store are different answers to different incidents.
  6. Responder wellbeing is a capacity decision, not a nicety. Rotation reduces effective team capacity on purpose.

Regulatory constraint that belongs in every playbook

AI Act Art 73(6): you may not alter the system in a way that affects subsequent cause evaluation before informing the competent authorities. Model rollback and memory purge are exactly such alterations. Notification goes in the first hour, alongside the 60-minute containment target. See serious incident reporting.

Reporting clocks

RegimeTriggerDeadline
Code of Practice C9Critical infrastructure disruption2 days
AI Act Art 73(3)Widespread infringement / Art 3(49)(b)2 days
Code of Practice C9Serious cybersecurity breach5 days
AI Act Art 73(4)Death10 days
AI Act Art 73(2)General15 days

Gaps

  • The CoSAI framework PDF itself is not in raw/ — only the announcement post. The CACAO playbooks, the five architecture patterns and the ATLAS mappings are the valuable parts and are unread.
  • Both previously-listed gaps are now closed: MITRE ATLAS (whose 63 case studies double as ready-made tabletop scenarios) and OWASP Top 10 for LLMs (which Microsoft cites for severity adaptation).
  • The Wiz source is half vendor marketing. The comparison table (conventional vs AI-driven IR) and the automated-vs-AI-driven distinction are the usable parts.

Log

See log.