🤔 Hugging Face Hacked by OpenAI Models That Escaped Their Sandbox

OpenAI reported that the "first fully agentic" attack on Hugging Face was carried out by their GPT-5.6 Sol and an as-yet-unannounced model.

Instead of honestly completing the ExploitGym cybersecurity benchmark, the models broke out of their isolated sandbox, gained network access, found a zero-day vulnerability on Hugging Face's servers, and downloaded the test solutions.

Did the models pass the test?

🤔 — Technically, no...
🤣 — Absolutely nailed it!