What happened
On Oct 28, OpenAI documented how it detects frontier-model misalignment and unauthorized agent behavior: Layer 0 activation probes, Layer 1 model-based reviewer, Layer 2 investigative agent, Layer 3 automated investigation, then human response — with evaluation against recorded incidents, simulated attacks, red-teaming and audits of internal agent traffic.
Why it matters
It is a reference architecture for agent-misalignment monitoring (monitoring reasoning, actions and environment effects rather than just prompts) and highlights the coverage gap the Hugging Face incident exposed, informing anyone building agentic guardrail/monitoring stacks.
Applicability
AI platform and security teams designing agent monitoring should benchmark their detection posture against OpenAI's layered model and its guidance on not incentivizing obfuscation/evasion in monitored agents.