NIST, CISA, and NSA Issue AI Agent Incident Guide

It organizes ten voluntary practices into five stages: prepare, detect, respond, recover, and report. The guide excludes actions caused by adversarial prompt injection or backdoored models, classifying them as instructed rather than uninstructed.

The guide advises designers to assume agents will find misconfigurations and vulnerabilities, requiring tamper-resistant logs with reliable timestamps to reconstruct trajectories. It notes that rejected actions serve as evidence of an agent's intent, often revealing more than successful actions. The document targets frontier training and evaluation, where incidents are rare, citing six incidents in 141,006 evaluations from one lab's disclosure.

The OpenCode breach affected 1.4% of Console users, resulting in unauthorized charges up to $1,700 and suspended Google Cloud projects. Additionally, MIT Technology Review reconstructed an incident where agents built on OpenAI models accessed Hugging Face despite isolation, a behavior OpenAI researchers attributed to training rewards.