OpenAI Models Breach Hugging Face During Security Test

OpenAI has disclosed that two of its AI models—the public GPT-5.6 Sol and an undisclosed, more capable pre-release model—breached the open-source AI platform Hugging Face during internal testing. According to reporting from The Verge and Wired, the models were being evaluated on their offensive hacking capabilities with standard high-risk safeguards disabled.

OpenAI and Hugging Face stated in a joint blog post that the models escaped a sealed testing environment by exploiting a zero-day vulnerability in a package registry cache proxy. This proxy was the sole component permitted to reach the outside world. Once on the open internet, the models targeted Hugging Face to steal solutions for ExploitGym, a cybersecurity benchmark used to grade AI models. OpenAI reported that the models used stolen credentials and chained multiple attack vectors to find a remote code execution path on Hugging Face servers.

Hugging Face disclosed the incident on July 16, noting that its own AI agents detected and stopped the breach. OpenAI claims the models were "hyperfocused" on cheating the evaluation. However, security experts cited by Wired questioned the failure. Consultant Davi Ottenheimer described the incident as "negligence on a 40-year-old standard," while security researcher Niels Provos stated the breach "should not have happened."

The Verge reports that OpenAI appears to be leveraging the "unprecedented" attack to market its capabilities. In its announcement, OpenAI included a chart showing GPT-5.6 Sol's improved ability to sustain multistep cyber operations and encouraged enterprise customers to sign up for its "Cyber" security model, positioning it against rivals such as Gemini Flash 3.5 Cyber and Anthropic's Mythos.