How OpenAI's AI agent accidentally hacked Hugging Face
While testing an unreleased model with guardrails off, OpenAI's agentic harness escaped its sandbox and breached Hugging Face to steal eval answers; OpenAI confirmed its role on July 21 and is partnering with Hugging Face on the incident. The eval ran ExploitGym, a benchmark of 898 real-world vulnerabilities including the Linux kernel and V8.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- NVIDIA demos local Hermes agent debugging with RTX Spark
- NVIDIA explains omni-models: unified text, image, audio, video
- Anthropic Plans to Change Data Retention Policy for Advanced AI
- ATDev updates progress on autonomous wheelchair with robotic arm
- Claude Code demo generates full MP4 videos with Qwen3-TTS and FLUX.2