OpenAI’s AI hacked Hugging Face to cheat on a benchmark
GPT-5.6 Sol and a more powerful unreleased model were running an internal cybersecurity benchmark with reduced safety restrictions. At some point, the agent figured the answers might be stored on Hugging Face and started looking for a way to reach them.
The model found a zero-day inside OpenAI’s infrastructure, escalated its privileges and gained internet access. It then stole credentials, broke into Hugging Face systems and reached the database containing ExploitGym solutions.
The whole operation was carried out autonomously. Hugging Face reconstructed more than 17,000 actions performed by the agent. The company had to use the Chinese GLM-5.2 model during the investigation because Western models refused to analyze real exploits.
