Skip to content

Back to the catalog

openai.comarticle

How we detect frontier model misalignment

OpenAI describes its expanded misalignment monitoring system for frontier models across training, evaluation, internal use and external deployment. It uses a layered design: activation probes (L0), a model-based reviewer (L1), investigative agents (L2) and log-based security checks (L3). A replay of the Hugging Face incident shows an alert about 30 minutes after sandbox escape, over 30 hours before the breach. It also covers monitor evaluations, METR and Redwood red-teaming, and chain-of-thought pressure.

TL;DR Explains OpenAI's layered monitoring for misaligned agent behavior, how it is evaluated and red-teamed, and what it caught.
Open original ↗
Also seen from another angle🛡️ Security →

Organisations named here

  • OpenAI
  • Hugging Face
  • METR
  • Redwood Research
  • OAI-repo Sabotage v2
  • Preparedness Framework
  • Sneaky Sandbox Escape

Technologies named here

  • GPT-6 Astra
  • Codex
  • GPT-5.5 Thinking
  • Responses API
  • Chat Completions API
  • GPT-5.6 Sol
  • GPT-5.6 Terra
  • Activation probes
  • GPT-5.6 Luna
  • OpenAI API
  • ChatGPT
  • LLM
  • gpt-5.5
  • Alignment investigator
  • Security investigator
  • Webhook alerts

Techniques named here

  • Agent red-teaming of monitors
  • Multilayer misalignment monitoring
  • Not training against monitor flags
  • Random sampling of unflagged traffic
  • Replay of recorded incidents
  • Simulated attack evaluations
  • Synthetic transcripts

Related in ai-safety