NIST publishes IR 8621, flagging gaps in AI incident management

NIST published Internal Report (IR) 8621, titled NIST Workshop on SI Incident Management, on November 2, 2026.

The report synthesizes findings from a May 14, 2026 workshop examining how organizations prepare for, detect and respond to incidents involving AI systems, which NIST now refers to under federal direction as super intelligence (SI).

What it is

The report examines the unique challenges of managing super intelligence (SI) incidents, including cybersecurity attacks that target SI systems as well as harms that arise from misuse, malfunction, or unintended actions.

It frames SI incidents into two overlapping groups: cybersecurity attacks that target SI systems and incidents that stem from SI misuse, anomalous behavior, or severe operational malfunctions resulting in harm.

What it requires

IR 8621 is a findings document rather than a binding rule, and it states that the report reflects stakeholder insights on current incident management practices, existing guidelines and resources, gaps in preparedness and response, and the roles and responsibilities of organizations involved in managing SI incidents.

Key areas for future work include updating NIST cybersecurity incident response guidelines, including SP 800-61, and exploring new resources to support incident readiness among model developers.

Additional activities include monitoring and evidence collection, incident information sharing, and clarifying coordination and responsibilities across model developers, deployers, and users.

Workshop participants also emphasized critical operational needs, including reasoning-layer telemetry, non-human identity governance, authorization-scoped containment, clear evidentiary and reporting schemas, and coordinated responsibility across distributed SI value chains.

What security leaders should do

Security teams should read IR 8621 as an early marker of where NIST's incident response references, including SP 800-61, are likely to be updated for AI-specific scenarios, and track that work as it progresses. Leaders should map which of their AI deployments could surface as a cybersecurity-incident scenario versus a misuse-or-malfunction scenario under the report's two-category framing, since each may call for different detection and escalation paths. Teams should also confirm they can capture reasoning-layer telemetry and non-human identity records for any AI agent involved in an incident, since the report lists these as gaps workshop participants raised. Finally, because the report calls for clearer coordination across model developers, deployers and users, security and legal teams should document in contracts and runbooks who is responsible for detection, containment, evidence preservation and reporting when an AI system behaves unexpectedly.