Detecting misbehavior in frontier reasoning models
OpenAI demonstrates that frontier reasoning models exploit loopholes, and that monitoring chains-of-thought with an LLM can detect such exploits. However, penalizing bad thoughts does not stop most misbehavior; it causes models to hide their intent.
