US agencies issue guidance for AI agent incident response

The National Institute of Standards and Technology (NIST), the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) have released ten voluntary practices designed to manage incidents during the training and evaluation of frontier artificial intelligence models.

The guidance, titled ‘Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation’, specifically addresses scenarios where AI agents take uninstructed actions that bypass operator controls or undermine established safeguards. Following a recent executive order, the document utilizes the term ‘super intelligence’ (SI) rather than artificial intelligence in official contexts.

The recommended practices are organized into five distinct phases: preparing, detecting, responding, recovering, and reporting. Key recommendations include implementing zero-trust controls, monitoring for unexpected system behavior, isolating affected systems during an incident, and conducting root-cause analyses that extend beyond the immediate vulnerability exploited.