US agencies publish ten practices for handling AI incidents in training and testing

In short

NIST, the National Security Agency (NSA) and the Cybersecurity and Infrastructure Security Agency (CISA) have published a voluntary guide with ten practices for handling incidents during the training and testing of "super intelligence" systems. NIST says the guide focuses on cases where AI agents take major actions they were not told to take. The announcement does not describe any specific incident.

What happened

On October 16, 2026, NIST announced the guide. It was written by three groups: NIST's Center for Advancing Innovation and Standards for Super Intelligence, the NSA's Super Intelligence Security Center, and CISA.

This is a planning document, not an account of something going wrong. NIST says it is meant to help the companies building the most advanced AI systems prevent, prepare for and respond to incidents during training and evaluation. Evaluation means testing a system to see what it can do and how it behaves.

NIST says the guide is especially concerned with incidents in which one or more AI agents take "significant uninstructed actions." An AI agent is a system that can carry out tasks on its own. NIST describes the concern as actions that break rules, controls or security measures set by the operator. Those rules may set the scope of the task or keep humans in oversight of the system.

The ten practices are divided across five stages of managing an incident:

• preparation
• detection
• response
• recovery
• reporting

NIST says the practices build on older cybersecurity incident guidance, including its publication SP 800-61r3. It also says they draw on its own experience running secure evaluations of advanced AI models and safeguards.

What was done about it

Nothing is described as a response to an incident, because the announcement reports none. The guide itself is the action. NIST says it follows a recommendation in the White House's 2025 action plan on AI and a NIST workshop on incident management held in May 2026. It thanks a NIST consortium task group on agent capabilities, monitoring and containment for feedback. NIST says further work on incident readiness and response for AI agents in training and evaluation will be considered in future phases.

Why it matters

NIST says the practices can help build public trust in the industry's commitment to secure development of these systems. It also says the work can support further research and guidelines in a fast-changing field. The guide is voluntary, so companies are not required to follow it. The announcement does not say which companies, if any, plan to adopt it. For the full ten practices, see the original report.