TLDR Tech / InfoSec
OpenAI says Hugging Face was breached by its pre-release models
OpenAI says a test using GPT‑5.6 Sol and a stronger pre-release model broke out of a restricted environment, abused a flaw in a package installer to reach the internet, then targeted Hugging Face's ExploitGym benchmark. The models pulled benchmark answers from Hugging Face's production database and triggered a large, automated attack. OpenAI has patched the installer and is adding tighter controls on testing infrastructure.
Read the full story