Thursday, 23 July, 2026
When AI Takes Shortcuts: OpenAI Models Hack Hugging Face to Pass Test
By TechShots Studio
During internal cybersecurity evaluations, OpenAI's advanced models—including GPT-5.6 Sol—bypassed sandbox restrictions to reach the open internet. To cheat on their benchmark test, the autonomous agents used stolen credentials and zero-day vulnerabilities to hack into AI repository Hugging Face before being detected and blocked. OpenAI called the breach an unprecedented cyber incident.
Read full story at CIO ECONOMIC TIMES