Supporting case
Federal agencies publish incident-response practices for frontier AI testing
What happened
On October 16, NIST, the National Security Agency and CISA published ten voluntary practices for incidents during advanced AI training and evaluation. The agencies use “super intelligence” in the document’s title; that terminology does not establish that the systems involved possess a particular level of capability. The guidance addresses agents taking significant actions outside their assigned task or human controls. Agency announcement ↗.
The 14-page document ↗ recommends limiting agent permissions, monitoring activity, preserving evidence, shutting down affected runs, rebuilding compromised environments and checking controls before restarting. It recommends notifying affected parties, voluntarily supplying AI-specific information to government experts and facilitating independent review. It separately tells organizations to comply with any existing legal reporting duties. Publication does not create a new mandatory reporting law, give outside evaluators access rights or establish an independent adjudicator.
For The Race, public agencies are specifying what responsible testing and incident handling should involve. This adds substance to the evaluation-governance question, but the consequential rights remain largely with operators unless an obligation incorporates these practices.
Dates and implementation stage
Development: 2026-10-16
