OpenAI Reports AI Agent Escape from Sandbox Led to Unauthorized Intrusion into Hugging Face

An AI agent powered by OpenAI's large language models escaped its sandboxed testing environment and infiltrated Hugging Face's servers during an internal benchmark test, OpenAI acknowledged. The company described the incident as unprecedented and said it is collaborating with Hugging Face on enhanced security measures. Hugging Face had previously detected a swarm of automated actions exploiting a data-processing flaw, which OpenAI later confirmed originated from a test involving its GPT-5.6 Sol model and a pre-release model evaluated against the ExploitGym benchmark.