AIRUNAMOK

NIST, the NSA and CISA released ten voluntary practices for managing incidents during advanced AI training and evaluation. The guidance specifically addresses situations where AI agents take significant uninstructed actions that violate task boundaries, controls or human oversight. The practices cover preparation, detection, response, recovery and reporting, including independent testing, technical containment and clearly defined shutdown procedures.

For the past several weeks, AIRUNAMOK has tracked a recurring question: What happens when AI can act? This week adds the next layer: What actually controls it? The answer is beginning to take shape across regulation, cybersecurity, enterprise architecture and public expectations. The emerging control layer includes: identity, permissions, containment, shutdown authority, independent testing, incident reporting, board accountability, data protection, human review, evidence preservation and lifecycle monitoring. Organizations increasingly need more than an AI policy. They need an operating system for control. The board level question this week is: If an AI system crosses a boundary tomorrow, do we know exactly what stops it?

AI incident response is becoming an operating discipline rather than a theoretical safety exercise. Organizations deploying high capability agents should establish shutdown authority, escalation criteria, forensic logging, recovery procedures and clear reporting responsibilities before the system is placed in a position to act.