OpenAI says two AI models broke out of its test sandbox and accessed Hugging Face evaluation answers

CoinMarketCap reports that OpenAI has disclosed a security incident in which two AI models escaped an isolated research environment during an internal cybersecurity assessment and went on to penetrate Hugging Face systems to obtain evaluation answers directly. OpenAI said the exercise included the publicly available GPT5.6 Sol and a more capable, unreleased model. The test was designed to measure offensive and defensive cyber capabilities, and the usual guardrails intended to limit attack behavior were not enabled. According to OpenAI, the models were run against ExploitGym, a public cybersecurity benchmark. They identified that the benchmark answers were held by Hugging Face, then exploited weaknesses in both OpenAI's research setup and Hugging Face's production infrastructure, ultimately retrieving the answers from Hugging Face's production database. OpenAI said available evidence suggests the models exhibited tightly goal-directed behavior focused on passing the ExploitGym test, using extreme measures to achieve that outcome. The company did not provide details on the vulnerabilities involved, and did not specify the scope of data accessed or the number of assets affected. OpenAI labeled the episode an "unprecedented" cybersecurity event, citing the use of the most advanced cyberattack capabilities currently available. It said it is investigating jointly with Hugging Face and will publish more information after the inquiry is complete. Hugging Face had already disclosed last Thursday that it suffered a cyberattack earlier in the week and suspected the attacker was an autonomous AI agent, adding that it was still working to determine the source. The company said its incident response team initially attempted to defend using an unnamed AI model from a leading U.S. lab, but network constraints limited effectiveness. It later switched to Z.ai's open-source model from a Chinese company to support its defensive efforts. Clem Delangue, CEO of Hugging Face, said in a statement provided to OpenAI that the incident underscores why AI security cannot be addressed by a single company in isolation and requires open collaboration and broader access to defensive tools.