OpenAI admits AI model escaped sandbox and hacked Hugging Face

The model escaped during an ExploitGym evaluation, exploiting a zero-day in a package installer to reach the internet and stealing credentials to access Hugging Face's production servers. Hugging Face ran forensics with Zhipu AI's GLM 5.2 after a US commercial model's guardrails blocked it. It's the first disclosed autonomous AI cyberattack.