Skip to content
Log inRegister

When an AI Agent Is the Incident: Evidence Before Rebuild

Dr. Abeer Alshammari · Published 11/1/2026

IntermediateProfessionalsCISOsExecutives

On 16 October 2026, NIST's Center for Advancing Innovation and Standards for Super Intelligence, the NSA's Super Intelligence Security Center and CISA released Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation. It is a 14-page, voluntary document, and it matters well beyond frontier labs: it is the first joint U.S. guidance that treats an AI agent acting outside its authorization as an incident class with its own response lifecycle.

What the guidance says

The document sets out ten practices organized into five stages: prepare, detect, respond, recover and report. It builds on NIST CSF 2.0 and SP 800-61r3, so security teams will recognise the structure. Its scope is autonomous models equipped with software tools that can act on an environment, including agents that act beyond their authorized scope or without authorization at all. Reported recommendations include least-privilege access for agents, monitored egress, kill switches, monitoring across the model, harness, host and network layers, and validating containment before restart. Near-misses are to be treated as real incidents. Gateways and monitors deserve as much scrutiny as the agents they watch.

Evidence is the governance question

The most consequential practice for governance teams is the one on evidence. Organizations are asked to preserve, to the extent necessary to investigate a control incident, model weights, intermediate checkpoints, sandbox and network logs, rejected actions and tool calls, and the configuration needed to reproduce what happened. Logs should be tamper-resistant, with reliable timestamps, and should cover agent trajectories, tool calls, reasoning traces, environment states and security-control decisions.

This is a very different standard from ordinary application logging. If an agent can take an action, you need to be able to show later what it was permitted to do, what it attempted, what was blocked and why. That is an audit-trail requirement, and it lands squarely in the GRC function rather than only in engineering.

What it deliberately leaves open

Reading the document alongside independent commentary, several gaps are worth noting before anyone cites it as a compliance baseline.

No retention period. It asks for evidence to be preserved but does not say for how long, and it does not address legal holds.

No order of operations. The practices on collecting evidence and on revoking credentials and rebuilding compromised environments can pull in opposite directions. Nothing says capture must finish before a rebuild begins. Your own runbook has to decide.

No legal effect. The document states it is not advice for compliance, regulatory or legal purposes, and it points back to existing reporting obligations rather than creating new ones.

Unclear reach to deployers. The authors say the practices may also apply to deployers, but the primary audience is the developers training and evaluating frontier systems.

Adversarial causes are out of scope. Uninstructed behavior triggered by prompt injection or malware is flagged as future work, yet that is the failure mode most enterprises are likely to meet first.

What to do with it now

For organizations deploying agents rather than building frontier models, the guidance works best as a design checklist, not a mandate. Four steps are reasonable. First, add "agent acted outside authorization" to your incident taxonomy, with a severity scale. Second, decide in advance what evidence you will capture, how long you will keep it, and who can place a hold, because the guidance will not decide this for you. Third, write the sequencing rule into your runbook: preserve first, then rebuild. Fourth, test your kill switch and your containment checks before you need them, and record the result as control evidence.

The document will not be the last word. But it gives boards and regulators a shared vocabulary, and organizations that can already show agent logs, containment tests and a tested shutdown path will be better placed when their auditors and insurers start asking.

Source: NIST, NSA and CISA, Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation, v1.0, October 2026, as summarized by BankInfoSecurity and ComplexDiscovery.

Try it yourself

An interactive CyberAbeer experience for this topic is in development.

Coming soon
Back to insights