NIST, NSA and CISA publish incident response guidance for AI agent control failures

What it is

This is a new voluntary joint guidance document, Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation, covering how organizations should handle incidents where autonomous AI agents take significant unauthorized actions. It was published on October 16, 2026 by NIST's Center for Advancing Innovation and Standards for Super Intelligence, the National Security Agency's Super Intelligence Security Center, and the Cybersecurity and Infrastructure Security Agency.

What it requires

The document targets what the agencies call an SI control incident, defined as one or more AI agents taking significant uninstructed actions that compromise the policies, controls, or security measures an operator imposed to scope a task or preserve human oversight.

The guidance organizes ten key practices across five stages of the incident management lifecycle covering preparation, detection, response, recovery and reporting.

The practices are grounded in existing cybersecurity incident management work, including NIST SP 800-61r3, rather than built as an entirely separate discipline.

CISA's topics page describes the same document as helping organizations that train and evaluate highly capable SI agents manage SI control incidents.

What security leaders should do

Teams running agentic AI training or evaluation environments should map the five lifecycle stages onto their existing incident response plans rather than building a parallel process from scratch. Because the guidance is voluntary, security leaders should treat adoption as a way to demonstrate readiness to regulators, customers and partners, and should ensure tabletop exercises and severity scoring specifically account for uninstructed-agent scenarios rather than only external-attacker scenarios. Given the explicit tie to NIST SP 800-61r3, aligning agent-incident playbooks with that existing framework should reduce duplication for teams that already maintain mature cyber incident response programs.