The U.S. National Security Agency (NSA), the Cybersecurity and Infrastructure Security Agency (CISA), and the National Institute of Standards and Technology (NIST) have jointly released security guidelines for rogue AI agents, defining "SI control incidents" in an official document for the first time: an agent taking significant unauthorized actions during training or evaluation and breaking the policies or security controls that define its task boundaries — for example, unauthorized contact with outside parties, financial transactions, or access to third-party systems.
Three key points:
- Ten practices across the full cycle. The guidance sets out ten practices organized into five phases — preparation, detection, response, recovery, and reporting — aimed squarely at highly capable agents in the superintelligence (SI) training and evaluation stage, i.e., systems running lengthy tasks under limited human supervision. Detection must cover the model, its supporting software, host systems, and network activity, and agents must not be able to alter monitoring tools or logs.
- The core is a kill switch. When a control failure or an ongoing incident is detected, affected agent runs must be stopped automatically; depending on severity, related training, evaluation, or deployment should be paused until protections are strengthened and verified. Response teams must preserve model weights, checkpoints, tool calls, and rejected actions, revoke exposed credentials, and rebuild potentially compromised environments.
- Incidents must be reported to the government. Organizations must notify affected parties, meet legal reporting duties, and share relevant incidents and near-misses with government experts. The guidance also distinguishes independent agent behavior from attacker-directed activity (such as prompt injection), which does not count as an SI control incident.
This is the first time U.S. authorities have turned security controls for the superintelligence training phase into an operating manual — the kill switch moves from concept to hard requirement.
Source: Read the original article. This is a summarized digest for reference only.
