Purpose
Extend the enterprise incident response (IR) program with AI-specific triggers, playbooks, and escalation paths so security, engineering, and compliance can rapidly contain AI misuse or failure.
Scope
Applies to any event involving AI systems that threatens confidentiality, integrity, availability, safety, compliance, or user trust, including vendor-provided AI incidents.
- Prompt injections causing data leakage
- Model hallucinations leading to incorrect decisions
- Unauthorized models deployed (“shadow AI”)
- Security breaches exploiting AI infrastructure
- Regulatory or contractual violations involving AI outputs
Definitions
- AI Incident: Event involving AI systems that requires coordinated investigation or remediation.
- Severity Matrix: Classification system mapping impact and urgency to response requirements.
- RCA: Root Cause Analysis documenting contributing factors and corrective actions.
Policy
All suspected AI incidents must be reported through the standard IR channel within 30 minutes of detection. Security Operations leads the response, with Model Owners, Responsible AI, and Compliance providing subject-matter expertise. Post-incident reviews and corrective actions are mandatory for Severity 1–2 events.
Roles and Responsibilities
Security Operations manages detection, triage, and communication. Model Owners supply technical details and remediation support. Responsible AI evaluates ethical/safety impact. Compliance handles regulator/customer notifications when required.
Procedures
Incident handling includes:
- Detection and triage using AI-specific alert rules and severity matrix.
- Containment measures (disable model, revoke tokens, switch to backup)
- Forensic data capture (prompts, logs, inputs/outputs)
- Stakeholder communication plan (executives, legal, customers, regulators)
- RCA documenting root causes, control gaps, and follow-up tasks
- Control updates (monitoring, guardrails, training) tracked to closure
Exceptions
If an incident turns out to be a false positive, Security Operations may close the ticket without RCA but must log rationale and detection tuning actions.
Review Cadence
IR metrics (MTTD, MTTR, number of Severity 1–2 incidents) are reviewed monthly. Playbooks are refreshed at least annually or after significant incidents.
References
- NIST AI RMF Manage/Govern functions
- ISO/IEC 27035 (Information security incident management)
- Internal documents: Incident Response Plan, AI Guardrail Playbook, Regulator Notification SOP