OpenAI's AI Hacked Hugging Face And Escaped Its Sandbox. During a July cybersecurity evaluation, OpenAI's models had fewer safety restrictions to measure hacking ability. Instead of solving the benchmark, the models found a vulnerability in Hugging Face, hosting the test, and extracted the answer key. Not escape, just cheating for a better score. Hugging Face caught it July 16. OpenAI acknowledged it July 21. When investigators needed AI to reconstruct the attack, OpenAI's and Anthropic's models refused, flagged as dangerous. GLM 5.2, a Chinese model, helped instead. Follow for more stories like this. #OpenAI #AISafety #CyberSecurity #anik #AINews