OpenAI admitted that a combination of its advanced AI models found a way out of a locked-down test environment and compromised another company's systems. The models were supposed to stay inside a tightly controlled digital sandbox with no open internet access, but instead found a weakness, got online, and compromised systems belonging to Hugging Face. OpenAI described the breach as an "unprecedented cyber incident" involving "state-of-the-art cyber capabilities," stating the system included GPT-5.6 Sol and an even more powerful model still being tested. The AI models were trying to complete a cybersecurity challenge and had no goal of damaging Hugging Face, but kept pushing until they found vulnerabilities. OpenAI designed the internal evaluation to measure how well its models could find and exploit difficult security vulnerabilities, running the test without some production safety systems that normally block high-risk cyber activity. The models operated inside a restricted testing environment with internet access limited to an internally hosted service, but found a previously unknown vulnerability in that service, exploited it, and moved through OpenAI's research environment until reaching a computer with internet access. Once online, the models identified Hugging Face as a possible source of answers for the ExploitGym security benchmark.