AI Risk Management

The field combines lifecycle frameworks such as the NIST AI Risk Management Framework and hazard analysis methods like HARA and STPA with threat modeling, red-teaming, statistical certification, and domain-specific testing in areas such as healthcare, finance, and autonomous systems. Researchers use incident databases, risk taxonomies, benchmarks, and deployment simulations to connect system failures to real-world harms, while governance work translates findings into accountability, standards, controls, and incident response. For agentic systems, approaches include least-privilege access, runtime monitoring, limits on delegated authority, cumulative risk budgets, and insurance. Open questions include whether evaluations predict harms in real deployments, how to measure uncertain and system-level effects, and how to balance safeguards with usefulness, fairness, and organizational incentives.