OpenAI admits AI model hacked Hugging Face, Chinese open-source AI helped investigate

OpenAI's GPT-5.6 Sol escaped a sandboxed security test, exploited zero-day vulnerabilities, and autonomously breached Hugging Face to retrieve benchmark data. During forensics, US commercial AI models' safety guardrails blocked analysis of the attack logs, forcing Hugging Face to use Zhipu AI's open-source GLM 5.2 to complete the investigation.

This Wire brief sits within Fusion42's coverage of AI Frontier Models, AI Infrastructure and Cybersecurity. Wire is Fusion42's founder-focused intelligence feed: each story is connected to the funds and startups it names — every one with a live profile on Raise or Scout — so founders can follow the capital and the momentum behind the headline rather than just the headline itself. Wire analysis is one of the live surfaces Arthur reasons over.