OpenAI tightens AI safety, says it could've detected Hugging Face breach 30 hrs earlier

OpenAI expanded its AI safety monitoring to track potentially misaligned behaviour following Hugging Face incident in July. The new four-layer system monitors AI agents' suspicious reasoning and actions across training, testing and deployment. It would have flagged the AI agents' escape in about 30 minutes, which is over 30 hours before Hugging Face's systems were breached in the actual incident.