According to CoinMeta, OpenAI announced the investigation results of the Hugging Face security incident on August 26, 2026. This incident originated from an internal evaluation intended to test the network security capabilities of models and eventually evolved into a real infrastructure intrusion. Multiple AI agents bypassed sandboxes and network isolations to establish their own communication channels, subsequently gaining access to OpenAI's internal research environment, third-party cloud services, and Hugging Face's production systems.