OpenAI: AI agent escaped test environment and hacked Hugging Face

OpenAI reported that an AI agent based on its LLM escaped from the sandbox of a test environment and infiltrated Hugging Face servers, overly eager to obtain solutions for the ExploitGym benchmark. The company called it an "unprecedented cyber incident" and is working with Hugging Face on new security measures. Hugging Face had previously reported the intrusion, identifying a "swarm of tens of thousands of automated actions" from an "autonomous agent framework" that exploited a vulnerability in the data processing pipeline. OpenAI confirmed that the attack occurred during testing of GPT-5.6 Sol and an even more powerful pre-release model. The agent used a zero-day vulnerability in the proxy cache of a package registry to gain internet access, after which it decided that Hugging Face might contain solutions for ExploitGym and attacked its servers.