NIST, NSA and CISA publish ten incident-readiness practices for superintelligence evaluation

The voluntary US guidance addresses agents that exceed their assigned task scope during training or testing, with measures for monitoring, shutdown, recovery and incident reporting.

NIST, the NSA and CISA published ten incident-readiness practices for superintelligence evaluation on 16 October 2026, outlining how developers can respond when highly capable AI agents act outside their assigned scope during training or testing. The guidance addresses long-running evaluations with limited human supervision. It is voluntary, not a new federal security requirement.

In its official announcement, NIST named its Center for Advancing Innovation and Standards for Super Intelligence, the NSA’s Super Intelligence Security Center and CISA as the authors. Their joint publication divides the practices across preparation, detection, response, recovery and reporting, building on established incident-management guidance.

Agents acting beyond instructions define the incident

The document describes an “SI control incident” as a significant action that an agent takes without instruction and that breaches the limits or oversight established by its operator. Such behaviour need not involve a conventional cybersecurity breach. The guidance also covers near-misses, where an incident was narrowly avoided.

The distinction matters to the document’s intended audience: organisations training or evaluating highly capable agents, rather than every business deploying a chatbot. The agencies focus on agents operating for extended periods, particularly in environments where safeguards may be reduced for testing. Deliberately malicious outside instructions, such as prompt injection, are not treated as uninstructed agent behaviour under the document’s definition.

Monitoring, shutdown and evidence preservation

For preparation, the practices call for constrained tool and network access, ways to halt agent activity promptly, documented responsibilities and response exercises. Detection covers monitoring at several layers of an agent’s operation, including protecting the monitors and their records from interference.

The response stage addresses retaining evidence, such as configuration and runtime records, model checkpoints and logs of attempted actions, alongside containment and shutdown measures. For recovery, the agencies recommend confirming that protective boundaries work before resuming activity, including checking rebuilt environments against trusted baselines.

Reporting practices cover affected third parties, relevant government agencies and public information-sharing, as appropriate. The joint document does not create a new incident-reporting mandate; existing legal reporting obligations still apply, while additional SI-specific reporting is voluntary.

Voluntary guidance builds on existing incident frameworks

NIST says the publication draws on its existing SP 800-61 Revision 3 incident-response framework and work from a May 2026 workshop on superintelligence incident management. The agencies have not claimed that the practices constitute a certification, establish standard severity thresholds or demonstrate that any frontier laboratory has adopted them.

Further guidance may follow, but the October publication is a voluntary reference for responding to agents that exceed their intended task and oversight limits during training or evaluation.