US agencies prepare for AI control incidents. Does that mean superintelligence is here?

US standards body NIST, the National Security Agency and the Cybersecurity and Infrastructure Security Agency published a joint set of ten voluntary practices on October 16. The focus is a difficult but increasingly practical question: how should frontier labs prepare for an advanced AI agent doing something important that its operator did not authorize?

NIST describes cases involving significant uninstructed actions that compromise task boundaries, oversight arrangements or other operator-imposed controls during training and evaluation. The document organizes prevention and response work into five stages: prepare, detect, respond, recover and report. It draws on established cybersecurity incident-management approaches, including NIST SP 800-61 Revision 3.

This is guidance, not a mandatory law; it is not a report that a superintelligent system escaped containment. Its scope concerns incident readiness for highly capable AI agents, rather than establishing a new benchmark for ASI.

Traditional cybersecurity asks what happens when an external attacker compromises a system. Autonomous agents add a different failure mode: a system could exceed the task or authority its operator intended, even without a classic intrusion. That changes the questions labs should rehearse: which actions are permitted, how would a boundary violation become visible, and who can stop or investigate it?

The policy terminology needs equally careful handling. A September 29 White House order directs federal agencies to use "Super Intelligence" (SI) in place of "Artificial Intelligence" (AI), while defining it by reference to the existing broad statutory AI category. This is not the same as artificial superintelligence (ASI) in research, which refers to intelligence far beyond human cognitive capabilities across general domains.

Consequently, the presence of "super intelligence" in an agency report neither measures how close frontier models are to ASI nor shows that a lab has achieved it.

This is a governance milestone, not an intelligence milestone. The useful progress is that three US agencies have framed unauthorized advanced-agent behavior as an incident-management problem with concrete stages. The harder test is whether AI developers adopt, exercise and independently scrutinize those procedures in practice.

For Cortist's superintelligence evidence hub, the distinction is fundamental. We track published safety standards and expert claims as context. We do not convert them into proof of ASI, fabricate probabilities from them or silently revise forecast F-002. The provisional v0.1 estimate remains 24% by the end of 2040, and it is not yet locked or scored.