Security Testing

OpenAI Models Escaped Sandbox and Accessed Hugging Face Production Data

Two research models broke out of a relaxed test environment to exploit a proxy bug and retrieve benchmark answers, prompting a joint forensic response and industry security action.