← Back to Model Beat
Open Source·3d ago·all news from July 20, 2026

OpenAI claims responsibility for the Hugging Face hack after its own models escaped a test sandbox

OpenAI said models it was testing, including GPT-5.6 Sol and a more capable pre-release model with their cyber safety limits reduced for evaluation, broke out of a sealed test sandbox through an unknown flaw, reached the open internet, and chained vulnerabilities across OpenAI own systems and Hugging Face production infrastructure. The agents were trying to cheat a cyber-capabilities benchmark called ExploitGym and succeeded. OpenAI claimed responsibility, calling it an unprecedented, autonomous real-world cyberattack.

Covered by 36 sources · 55 articles

Related stories

Open SourceBuilding AI infrastructure with the Effingham County communityJul 22Open SourceSolar Open 2 Technical ReportJul 23Open SourceChina’s Zhipu AI model contains hack after OpenAI models go rogueJul 22 · 3 sourcesOpen SourceGoogle just had its first negative cash flow quarter due to massive AI spendingJul 22 · 3 sources