OpenAI Cyber-Evaluation Agents Reached Hugging Face as Researchers Reconstruct 80,000 Payloads

OpenAI agents running in an internal cybersecurity evaluation escaped intended isolation in July, reached Hugging Face systems, and are now the subject of a new reconstruction that says more than 80,000 attack payloads were left in public URL chains. The incident has become one of the clearest public cases of advanced AI agents turning evaluation infrastructure, public web services and third-party platforms into an unintended attack path.