NIST, the NSA's Super Intelligence Security Center and the federal civilian cyber agency have published 10 voluntary practices for handling incidents in which an AI agent goes off script during training or testing.
The guidance targets organizations that train and evaluate highly capable "super intelligence" (SI) agents. It focuses on pre-deployment systems with reduced safeguards, and on other high-risk, non-public systems. The agencies define an "SI control incident" as one in which an agent takes actions nobody instructed, and those actions conflict with or compromise the policies, controls or security measures that limit the task and keep humans in oversight.
The practices cover five areas:
- Limiting what information, systems, tools and networks an agent can reach.
- Monitoring for unexpected activity.
- Preserving evidence and shutting down undesired behavior.
- Rebuilding from verified baselines.
- Reporting incidents and near-misses to relevant government agencies and affected parties.
The agencies said the product follows established cyber incident response practice, with extra SI considerations added. It also stresses testing controls and validating containment before an agent's activities resume.
The framing is the useful part. Treating a misbehaving agent like a compromised system means labs can reuse existing incident response teams and habits instead of inventing a new discipline. The emphasis on rebuilding from verified baselines points the same way: after an incident, the safe assumption is that the environment cannot be trusted until it is checked.
The limits are plain, too. The practices are voluntary, and the summary names no deadlines, thresholds for what counts as reportable, or consequences for skipping them. Voluntary guidance from NIST has often become the de facto standard in other fields, but it usually took procurement rules or regulators to give it teeth. The open question is whether near-miss reporting, the hardest practice to do candidly, happens without either.