Federal AI incident guidance leaves rebuild sequencing open
For agents with credentials, the missing order of operations turns a recovery call into an evidence-retention problem.
TL;DR
MeriTalk’s Kiteworks-authored column flags a gap in the October 16, 2026 federal guidance for highly capable super intelligence agents. The practices separately tell responders to preserve agent evidence and to rebuild suspect environments after revoking touched credentials, but don’t set the order. The main audience is frontier training and evaluation teams; any organization giving agents credentials inherits the audit problem.
The official document is careful, which is why the gap matters. CISA’s summary says the October 16, 2026 product was issued with the National Institute of Standards and Technology and the National Security Agency’s Super Intelligence Security Center as ten voluntary practices for organizations training or evaluating highly capable super intelligence agents, especially pre-deployment systems with reduced safeguards or other high-risk non-public systems. CISA puts evidence preservation next to undesired-behavior shutdown and rebuilding from verified baselines. It doesn’t say which move has priority when the same environment is on fire.
MeriTalk’s opinion piece by Frank Balonis, Kiteworks’ chief information security officer and senior vice president of operations, presses on that seam. The source article says Practice 5 asks teams to preserve model weights, sandbox and network logs, rejected tool calls and the agent’s reasoning traces; Practice 6 asks them to revoke credentials the agent touched and rebuild environments that may be compromised; Practice 7 adds verification and containment before agents resume activity. Those are reasonable steps. They just aren’t operationally ordered.
The guidance can be reasonable and still leave a compliance hole. Voluntary federal practices often avoid dictating legal holds or retention periods. But agent incidents make the evidentiary object stranger: the record may need to prove both instruction scope and permission state, along with the controls touched. If the rebuild destroys that record, the post-incident story becomes confidence in place of evidence.
NIST’s November workshop summary separately says participants identified limits in traditional response lifecycles when applied to SI systems with autonomy and deep integration. That supports the modest lesson here. Before an organization hands an agent credentials, the incident plan needs a preservation checkpoint that survives the pressure to rebuild. Otherwise the first clean environment may come with the dirtiest audit trail.
Published ·Deep Fathom