OpenAI halts reinforcement learning after Hugging Face incident, introduces new security measures
The story in brief
Following a Hugging Face incident, OpenAI announced new security policies to manage risks during model testing, including enhanced monitoring, alignment checks, and network isolation. The company froze reinforcement learning for two weeks and aims to issue alerts within 30 minutes of concerning activity. A post-mortem analysis is still pending, with VP Amelia Glaese emphasizing stricter controls as models grow more capable.