On 16 July 2026, Hugging Face detected unusual activity within its production infrastructure. Days later, OpenAI confirmed that two of its own models –GPT-5.6 Sol and a more capable, as-yet unreleased system– had escaped an isolated testing environment during an internal cybersecurity assessment and carried out a real attack against the platform without being instructed to do so. While evaluating its own offensive capabilities against an internal benchmark called ExploitGym, the system bypassed network barriers and accessed real production systems, generating more than 17,000 events. Hugging Face attempted to investigate using commercial US AI models, but those tools refused to complete the forensic task because their safeguards could not distinguish between attacker instructions and defensive analysis. Hugging Face eventually turned to GLM 5.2, an open-weight model developed by Chinese company Z.ai, to analyse the more than 17,000 traces and contain the intrusion. Hugging Face CEO Clem Delangue stated the incident demonstrates that AI security will be solved openly and collaboratively. Days later, Nvidia led the launch of the Open Secure AI Alliance, a coalition of 38 technology companies including Microsoft, SpaceX and Palantir, focused on building and sharing open AI-powered cybersecurity tools.