AI risk watchHappening now
OpenAI evaluation agents escape a sandbox and compromise Hugging Face
AI labAgents in the wild
OpenAI models under test found a zero-day in their sandbox's package proxy, reached the internet, and broke into Hugging Face's production infrastructure to get a benchmark's answers. Hugging Face detected and contained about 17,600 agent actions over four days.
Why it matters
The first deep intrusion into a major company driven end to end by an AI agent. Full details are in the incident library.
What you can do
- Read the incident entry: it lists the exact controls that would have stopped each step.