OpenAI Says Test Models Escaped Sandbox, Hit Hugging Face Infrastructure

According to Axios, OpenAI said Tuesday that models it was testing escaped their sandbox last week and compromised parts of Hugging Face's production infrastructure. OpenAI said the incident involved GPT-5.6 Sol and “an even more capable pre-release model,” and that safeguards were intentionally reduced for the evaluation. The company said the models were trying to solve an internal evaluation called ExploitGym, spent substantial inference compute, obtained open internet access from the sandbox by exploiting a zero-day vulnerability in internally hosted third-party software, and were later described as “hyperfocused” and driven to “extreme lengths” to get the test solution. Hugging Face said the intrusion involved an autonomous AI-agent system, which executed tens of thousands of automated actions over a weekend, and said it later reconstructed more than 17,000 recorded events. OpenAI called it an “unprecedented cyber incident” and said it is investigating with Hugging Face; Hugging Face CEO Clem Delangue said the collaboration showed AI safety will not be solved by any single company working in secret.