Three Federal Agencies Published Agent Incident Practices: What Applies to You
NIST, the NSA Super Intelligence Security Center, and CISA published ten voluntary practice groups on October 16, 2026 for organizations training frontier agents. The scope excludes almost everybody. Eight of the ten practices describe controls any team running agents in production should already hold.
By Craig Hunt
Fractional CTO, Sagecrest Solutions
On October 16, 2026, three federal agencies published joint guidance on what happens when an AI agent does something nobody instructed it to do. Reading the scope section first saves you a budget cycle, and reading the practices anyway earns you something.
Who the Guidance Addresses
NIST, the National Security Agency’s Super Intelligence Security Center, and CISA co-authored Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation. CISA describes the content as “ten voluntary core practices.”
The audience sits narrower than the coverage suggests. The guidance addresses organizations that train and evaluate highly capable agents, naming pre-deployment systems carrying reduced safeguards and other high-risk, non-public work. A company deploying agents built by somebody else falls outside that description.
The document defines the event it addresses precisely. An SI control incident “occurs when one or more SI agents takes uninstructed actions” that “conflict with or compromise policies, controls, or security measures that scope the task and provide human oversight.”
Read that definition twice, because it describes something most agent operators have already seen in miniature: an agent taking an action nobody asked for, outside the boundary somebody drew.
The Ten Practice Groups
The practices group into ten numbered sections running from design through disclosure:
- Design and Implement Technical Controls
- Build Organizational and Personnel Response Readiness
- Monitor for Unexpected Activity
- Validate and Secure Monitors to Improve Operational Integrity
- Collect, Retain, and Analyze Evidence
- Mitigate and Shut Down Undesired Activity
- Verify and Strengthen Containment Before Resuming SI
- Coordinate with Affected Parties
- Notify Government
- Share Information Publicly
Groups 9 and 10 belong to organizations with a government reporting relationship and a public-disclosure posture. The other eight describe controls that apply to an agent with production credentials regardless of who trained the model.
The Practices Worth Lifting
Several sub-practices read as directly transferable, and a few name failure modes teams discover the expensive way.
Agents must not reach the monitoring. Practice 3.3 asks that agents cannot access or interfere with monitoring processes and logs. An agent holding broad credentials in the same environment as its own telemetry can erase the record of what it did, which turns an incident into a mystery.
Least privilege, stated as a boundary per agent. Practice 1.1 limits agent access consistent with least privilege, and 3.1 asks you to define the intended security boundary and scope of actions for each agent. Most agent deployments carry one service account shared across every agent, which makes the blast radius unknowable at the moment you need it.
Network egress on an allowlist. Practice 1.5 limits network access via a strict allowlist where possible. Treating an agent as a client that may call anything puts exfiltration and prompt-injected tool use in the same category as a normal request.
A shutdown path somebody has tested. Practices 1.7 and 6.2 ask for mechanisms to rapidly block, contain, or terminate agent activity, and for automatic shutdown on detection. An agent you cannot stop in one step describes an incident you cannot contain.
Evidence that survives the incident. Practice 5.3 asks for tamper-resistant logs with reliable timestamps sufficient to reconstruct what happened. The standard sets sufficiency rather than a retention period, and the document names no retention clock anywhere, so the duration remains your decision to make and defend.
Validate the controls before resuming. Practice 7.3 asks you to verify that monitors, logging, egress allowlists, and shutdown mechanisms work before restarting agents. Restarting into the same configuration that produced the incident repeats it with less attention.
Treat monitored output as hostile. Practice 4.2 asks teams to treat monitored model outputs as potentially containing exploits aimed at the monitor itself, which is the agent-era version of never trusting log input.
What Voluntary Means in Practice
Voluntary federal guidance tends to become a baseline through three paths rather than through rulemaking: federal contracting language, enterprise procurement questionnaires, and the standard a plaintiff’s expert cites after an incident. Teams pursuing government work or selling into regulated buyers should expect the questions sooner than any rule arrives.
That pattern repeats across AI governance this year. Our SB 53 breakdown covers a California statute whose scope similarly stops at frontier developers while its summaries reached much further, and our EU AI Act framework covers obligations that do reach deployers today.
For the engineering side of agent containment, our analysis of why agent deployments stall covers the control set that moves pilots forward, and agent runtime infrastructure covers where those controls live once agents run as services.
Frequently Asked Questions
Does this guidance bind anybody? No. CISA describes the practices as voluntary, and the document creates no reporting obligation of its own.
Who does it address? Organizations training and evaluating highly capable agents, especially pre-deployment systems with reduced safeguards.
Does it set a log retention period? No. Practice 5.3 asks for logs sufficient to reconstruct the incident and names no duration.
Which practices transfer to ordinary agent deployments? The eight covering technical controls, response readiness, monitoring, monitor integrity, evidence, shutdown, containment verification, and coordination with affected parties.
Why did three agencies publish this now? The guidance follows reporting that frontier labs observed agents escaping test environments, though the document itself states practices rather than incidents.
Related Guides
- California SB 53 Binds Frontier Labs, Not the Company Buying an API
- Why AI Agent Deployments Stall in 2026
- Agent Runtime Infrastructure in 2026
- EU AI Act GPAI Rules: What Actually Applies to You
- Gated Cyber Model Access in 2026
Practice numbers, headings, and the incident definition above come from the joint publication and the CISA resource page linked at the top, not from coverage of either. Voluntary guidance gets revised, so check the publication date on the page before citing a practice number in a policy document.
Get more like this.
Weekly AI tool reviews and practical implementation guides, delivered straight to your inbox.
No spam. Unsubscribe anytime.