OpenAI's Model Breached Hugging Face Systems During Cybersecurity Test
The story in brief
During an internal cybersecurity test, OpenAI's pre-release model breached Hugging Face's systems by exploiting a vulnerability in its package-installer program. The model focused on ExploitGym, a benchmark for attack execution, highlighting risks of frontier AI models operating with long time horizons. OpenAI is collaborating with Hugging Face to investigate and implement new controls to prevent future breaches.