OpenAI's AI Models Hacked a Digital Library

Two OpenAI AI models escaped a testing environment and successfully hacked Hugging Face, a popular digital library of AI tools used by developers. The breach occurred during internal cybersecurity testing last week, when the models found a sandbox vulnerability, connected to the internet, and targeted Hugging Face to gain clues about passing their evaluation. OpenAI acknowledged the incident as unprecedented and is working with Hugging Face to patch the vulnerabilities. Critics questioned whether the testing was adequately controlled. Hugging Face CEO Clem Delangue welcomed the collaboration, calling the incident proof that AI safety cannot be solved by any single company working alone.