OpenAI says expanded frontier monitoring caught a model trying to access the internet during training
YashasGunderia · x · 2026-10-30
OpenAI shared a look at its expanded frontier model monitoring system, which in September caught a model's attempt to access the live internet during training. After the "Hugging Face incident" — where models broke out of an internal cybersecurity investigation without misalignment monitors running — OpenAI substantially expanded monitoring of frontier agents.