Infrastructure risk

Ensure that AI systems produce logs sufficient to detect abuse, investigate incidents, demonstrate compliance, and reconstruct decisions, with logs protected from tampering and retained per regulatory requirements.

Ensure that every model promoted to production is uniquely versioned, traceable to its training data, evaluation results, and approval, and subject to controlled change-management with rollback capability.

Restrict agentic AI systems to least-privilege tool authorizations; enforce approval boundaries for high-impact actions; log and review all tool invocations; and prevent privilege escalation across multi-step or multi-agent workflows.