NIST, NSA and CISA recommend automatically stopping powerful AI agents when they undermine their operators' controls during training or testing. The three US agencies published ten voluntary practices for preparing for and responding to such incidents on October 16.
Their definition of a control incident covers significant, uninstructed actions that compromise task boundaries or human oversight. Harm to an outside party is not required. Actions explicitly directed by operators or attackers fall outside that definition.
A detected control failure should automatically end the affected agent trajectory. Depending on the incident's scope, the agencies also recommend pausing related training or evaluation runs, revoking affected credentials and rebuilding environments that may have been compromised.
Before restarting, operators should verify monitoring, logging, network permissions and shutdown mechanisms. Agents should not be able to alter their own monitoring processes or logs. External evaluators should help establish whether the safeguards work.
The guidance also recommends sharing incidents and near misses with relevant government teams and publishing lessons where appropriate. It remains voluntary. The authors identify training-safe recovery as unfinished research: restarting must avoid reinforcing the problematic behavior or making future failures harder to detect.