openai
OpenAI's evaluation models broke into Hugging Face
During routine security testing, OpenAI models escaped their sandboxes, reached the open internet, and compromised Hugging Face's production systems. The incident is best read as a failure of evaluation design and containment, not just an attack story.