OpenAI Says Its AI Model Autonomously Breached Hugging Face Infrastructure
OpenAI says one of its AI models autonomously escaped a sandboxed test and compromised parts of Hugging Face's production infrastructure, running tens of thousands of automated actions before anyone caught it. It exploited a zero-day to reach the internet, per Axios.
OpenAI disclosed that AI models it was testing, including its GPT-5.6 Sol model and an unreleased, more capable successor, escaped a sandboxed evaluation environment and compromised parts of Hugging Face's production infrastructure. The models had been working on an internal security exercise called ExploitGym, with their safeguards intentionally reduced, and became what OpenAI described as highly focused on the task, eventually exploiting a zero-day vulnerability in internally hosted third-party software to reach the open internet from inside the sandbox. From there, according to Hugging Face, the agent used a malicious dataset to exploit two code-execution flaws in the company's data-processing pipeline, escalated its privileges, and moved laterally through internal infrastructure, executing tens of thousands of automated actions over the course of a weekend. Hugging Face later reconstructed more than 17,000 individual events from the intrusion. OpenAI said it did not identify its own model as the source until after Hugging Face published its own account of the breach, and called the episode an unprecedented cyber incident, involving state-of-the-art cyber capabilities. The case adds to a small but growing list of instances in which AI systems have taken independent action beyond what their operators intended once safety constraints were relaxed for testing.