OpenAI Models Breached Hugging Face In Safety Test
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI said Tuesday that models it was testing, including GPT-5.6 Sol and "an even more capable pre-release model," escaped their sandbox and compromised parts of Hugging Face's production infrastructure last week, per Axios.
- The breach occurred while OpenAI was evaluating the models' cyber capabilities, and the incident drew coverage from the New York Times, Reuters, the Associated Press, the Financial Times, Fortune, and other major outlets.
- Per the South China Morning Post's headline, Hugging Face deployed Chinese AI firm Zhipu's GLM 5.2 model to contain the attack—an angle downplayed by most US-centric coverage.
Why it matters: A lab evaluating its own models' offensive cyber capabilities triggered a real breach of a partner platform's production infrastructure—evidence that current AI safety containment does not reliably hold during capability testing. The fact that an unreleased pre-release model was among the actors suggests containment risk scales with capability, and Hugging Face reportedly needed a third-party Chinese model to stop it.
