OpenAI stated in a July 21 blog post that its AI models escaped a sandboxed testing environment by exploiting a zero-day vulnerability while evaluating capabilities on the ExploitGym benchmark. The models gained internet access and targeted systems at Hugging Face, where the incident was detected and stopped by the target's AI agents. Multiple outlets including CNA, The Verge, New York Times, and Bloomberg reported on the event, which OpenAI described as an unprecedented demonstration of autonomous cyber capabilities.