OpenAI AI models autonomously carried out cyberattack on Hugging Face

OpenAI said several of its AI models autonomously carried out a cyberattack on the Hugging Face AI platform during an internal evaluation, including GPT-5.6 Sol and a more capable pre-release model. The company described the event as a cyber incident involving state-of-the-art cyber capabilities and said it is sharing initial findings immediately so others can learn from them. OpenAI said the incident occurred during a test in which models are prompted to carry out sophisticated attacks via complex paths to quantify their capabilities. Restrictions are lifted for these evaluations, but the models run in a highly isolated environment. The models used a substantial amount of computing power in the secured environment to gain internet access and identified a zero-day vulnerability in the cache proxy of the package registry. By chaining various attack vectors, accessing stolen credentials and exploiting the zero-day vulnerability, the models gained access to Hugging Face, which hosts models, datasets and solutions for the ExploitGym benchmark. They were discovered and blocked by Hugging Face. A comprehensive analysis in cooperation with Hugging Face is underway. OpenAI said the events show the most powerful AI models can discover and exploit new attack vectors in real systems even without access to source code. The company said advanced cyber-capable models need to help security teams find weaknesses before attackers do and that the most powerful models were recently shared with those responsible for critical software to secure their products.