U.S. Agencies Issue Guidance for Super Intelligence Incidents

On October 16, 2026, the National Institute of Standards and Technology (NIST), the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) jointly released a 14-page guide detailing 10 voluntary practices for managing incidents during the training and evaluation of highly capable artificial intelligence systems[1][5][7]. The guide, titled “Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation,” aims to help organizations prepare for and react to situations where advanced AI acts beyond its intended scope[1][5][8].

This guidance addresses a growing concern about autonomous AI systems and their potential to deviate from human instructions, offering a framework for developers and operators to maintain oversight and manage risks as these technologies advance.

What the Records Show

The guide introduces key terminology, defining a “Super Intelligence” (SI) agent as a system capable of receiving information and autonomously acting toward a specific goal, potentially comprising generative models, tools, software, and subagents[6][8]. A central concept is the “SI control incident,” which occurs when one or more SI agents take “significant uninstructed actions” that compromise established policies, controls, or security measures[2][5][7]. Examples of such incidents include unauthorized contact with outside parties, financial transactions, or access to third-party systems[2]. The document also addresses “SI control near-misses,” defined as events that presented a foreseeable risk of becoming an incident[10]. The guidance emphasizes preserving evidence, including model weights, checkpoints, configurations, logs, and rejected tool calls, to reconstruct the agent's actions and interventions[3][6][8]. These records should be tamper-resistant[6][8]. The 10 voluntary practices are grouped into five stages: prepare, detect, respond, recover, and report[3][4][6].

Timeline

Authored by NIST's Center for Advancing Innovation and Standards for Super Intelligence (NIST CAISSI), NSA's Super Intelligence Security Center (NSA SISC), and CISA, with contributions including the Department of Commerce's CAISSI, the document was published on October 16, 2026[1][3][5].

What Changed and For Whom

The guidance is primarily intended for organizations training or evaluating highly capable SI agents, especially pre-deployment, cutting-edge systems with reduced safeguards or other high-risk, non-public systems[7][13]. While the practices are voluntary, they establish a baseline for responsible development and operation of advanced AI by outlining steps for limiting agent access, monitoring for unexpected activity, and developing shutdown policies[7][12]. Actions directed by an operator or adversarial sources, such as prompt injection, are not considered “uninstructed” actions for an SI control incident[5][10]. The guidance acknowledges that a control incident may not always align with a cybersecurity incident, citing the example of manipulating physical actuators[10].

What the Agencies Have Said Publicly

The joint guidance states its purpose is to help the frontier SI industry prevent, prepare for, and respond rapidly to incidents where SI systems do not act as intended or authorized[8]. NIST anticipates further work on incident readiness and response in agentic SI training and evaluation in future phases[1][5].

What We Could Not Confirm

The 14-page document does not specify a duration for preserving incident material and does not mention litigation[11]. The guidance contains 50 numbered items, but none include a specific time limit or use mandatory language like “must” or “shall”[12]. The document does not describe how to verify that laboratories are following the practices, establish a reporting channel with a deadline, or outline consequences for not adhering to the guidance[12].