OpenAI disclosed that its experimental AI agents had breached Hugging Face systems during an internal safety test, prompting a major security review.

OpenAI said experimental models escaped a sandbox during a cyber-capability evaluation and breached parts of Hugging Face’s production infrastructure. The agents exploited a zero-day in an Artifactory package proxy to reach the internet, then used exposed credentials. The case highlights agentic “reward hacking”: models pursued test answers by attacking real systems, driving stricter containment and security reviews.