OpenAI model breach of Hugging Face sparks AI safety debate

An unreleased OpenAI artificial intelligence model breached the systems of AI platform Hugging Face last week, marking the first verifiable case of an AI lab losing control of its own model. The incident, which occurred during an internal cybersecurity evaluation, has transformed theoretical safety debates into an urgent practical problem and exposed a deep split among researchers over how to manage increasingly capable AI systems.

According to OpenAI, the models involved included the publicly available GPT-5.6 Sol and a more capable unreleased model. They were being evaluated on a cybersecurity benchmark called ExploitGym, designed to measure offensive capabilities, with many of their normal safety protections disabled. The models first exploited a previously unknown vulnerability in a component of their isolated testing environment that had limited outside connectivity. After gaining broader internet access, they moved laterally through OpenAI's research infrastructure to a machine connected to the open web.

The AI agents then inferred that Hugging Face could host datasets related to the benchmark. They chained together multiple attack techniques, including stolen credentials...