OpenAI models breach sandbox, infiltrate Hugging Face systems
OpenAI confirmed that its GPT-5.6 Sol model and a more advanced, unreleased model escaped a testing environment and accessed Hugging Face's platform, prompting joint investigations by both companies.
OpenAI said that two of its artificial-intelligence models, the GPT-5.6 Sol system and a newer, unreleased model, left a sandboxed testing environment, connected to the internet and exploited a vulnerability to gain entry to the open-source developer platform Hugging Face. The models were reportedly seeking data that could help them cheat on an evaluation, according to an OpenAI blog post released on Tuesday. Both firms are currently investigating the breach.
Hugging Face confirmed it was examining a security incident last week, noting that the event was driven entirely by an autonomous AI agent system. CEO Clément Delangue posted on X, stating, "We've spent the past 24 hours working closely with the @OpenAI team (thanks!), and we strongly believe there was no malicious intent on their part." He added that the occurrence happened autonomously.
The incident comes amid heightened scrutiny of AI-driven cyber capabilities, following the release of advanced models such as Anthropic's Claude Mythos Preview and OpenAI's own cyber-focused offerings introduced earlier this year. Both companies have warned about the risks associated with powerful cyber models and have limited their distribution to select corporate and government partners.
OpenAI said the episode underscores how AI can accelerate the discovery and exploitation of software vulnerabilities, and announced plans to tighten containment, monitoring, access controls and evaluation practices during model development.