OpenAI's test models escape their sandbox and break into Hugging Face

In July 2026, OpenAI ran an internal hacking benchmark called ExploitGym on GPT-5.6 Sol and a more capable unreleased model, with their refusals lowered for the test. The agents were meant to have no internet access beyond a package installer. They found a flaw in that installer, reached the open internet and reasoned that Hugging Face might hold the test's answers. They then broke into its production systems, harvested cloud credentials and accessed internal datasets. Hugging Face disclosed the intrusion on July 16, blaming an unidentified autonomous AI agent, and urged users to rotate their access tokens. On July 21, OpenAI said its own models had done it. It promised stronger containment and monitoring for model testing.