NIST and CISA Release Framework for AI Agent Incident Response

The National Institute of Standards and Technology (NIST), the Cybersecurity and Infrastructure Security Agency (CISA), and the National Security Agency (NSA) have released joint guidance for managing security incidents involving autonomous AI agents. The document, titled “Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation”, establishes identity records as the primary evidence when agents act without human direction.

The 14-page guide is designed for organisations training or evaluating highly capable AI agents that operate over long durations with limited supervision. It defines a “control incident” as an event where an agent takes significant actions that undermine policies, controls, or human oversight, regardless of whether an external adversary is involved.

A central theme of the guidance is the shift from investigating intent to investigating identity. While traditional incident response often focuses on why an event occurred, the agencies argue that the first priority in an AI incident must be determining whose credentials were used. This approach reflects the growing complexity of autonomous AI hacking teams and agents capable of executing multi-step tasks.

Identity as the Primary Forensic Evidence

The guide organises its recommendations into five stages: prepare, detect, respond, recover, and report. It emphasises that agent incident readiness depends on maintaining a clear record of which identity acted and what specific authorities were granted to it before the incident occurred.

Recent incidents have demonstrated the risks of poorly managed agent credentials. In October 2026, the OpenCode platform reported a breach where API keys for external providers including OpenAI and Anthropic were exposed. The incident affected approximately 1.4% of users and resulted in unauthorised charges and suspended cloud projects. Responders faced challenges because keys stored by agent platforms are often only revocable by the account owner at the issuing provider.

The NIST and CISA guidance suggests that organisations must maintain a comprehensive inventory of every credential an agent can access. Without this record, it is impossible to effectively revoke access during the “respond” phase of an incident. This is a critical component of a broader data breach response checklist for modern enterprises.

Managing Session Accountability

The agencies also recommend the implementation of tamper-resistant logs that can reconstruct agent trajectories. This concept, known as session accountability, requires distinguishing between actions taken by a human and those taken by an agent acting on their behalf.

Research indicates the scale of this challenge. Telemetry samples from coding agents show they can request as many as ten shell commands for every single human prompt. If logs cannot separate these actions, investigators cannot accurately determine the source of an unauthorised command. The guidance encourages designers to use single-use infrastructure and credentials that expire as soon as a task is completed.

Organisations are advised to measure their readiness by how quickly they can enumerate the credentials available to a specific agent identity. If these answers take days rather than minutes to produce, the guidance suggests the organisation is unprepared for the rapid tempo of an autonomous AI agent security incident.