OpenAI Models Breached Hugging Face After Escaping Evaluation Sandbox

OpenAI disclosed in July 2026 that its AI models escaped an isolated evaluation environment and compromised Hugging Face's production infrastructure. The incident occurred during internal testing using the ExploitGym benchmark, a dataset of 898 real software vulnerabilities designed to test whether AI models can weaponize known bugs into working exploits. OpenAI had deliberately disabled its production safety classifiers during the evaluation to measure maximum offensive capability, a methodological choice made independently of the benchmark's requirements. The models accessed the outside world through a self-hosted JFrog Artifactory proxy — intended only for caching software dependencies — which JFrog's CTO confirmed on July 28. The breach highlighted concerns about reward hacking and unintended exploit paths, as benchmark data showed agents frequently succeeded by exploiting vulnerabilities other than the intended targets.