Prepare for AI incidents before deployment.

Advanced AI governance needs explicit detection, containment, escalation, evidence preservation, recovery and learning mechanisms.

Identify abnormal behavior, harmful outcomes, control failures or material policy breaches.

Restrict access, disable tools, reduce permissions, isolate integrations or suspend deployment.

Route the incident by severity to technical, risk, legal, security and executive owners.

Secure logs, prompts, outputs, model/version metadata, decisions and affected integrations.

Remediate causes and require explicit approval before restoring higher-risk functionality.

Feed findings into evaluation suites, policies, thresholds and future deployment decisions.

Severity model

Severity should combine impact, scope, reversibility, control failure and uncertainty.

Level	Illustrative condition	Governance response
Level 1	Low-impact anomaly, contained, no material control breach.	Log, review, tune monitoring.
Level 2	Repeatable harmful behavior or limited control failure.	Restrict scope, investigate, owner sign-off.
Level 3	Material harm potential, privilege misuse, significant policy breach.	Containment, senior escalation, independent review.
Level 4	Severe or systemic risk, loss of effective control, broad impact.	Stop authority, executive response, external obligations assessment.