OpenAI AI Model Breached Sandbox, Accessed Hugging Face Infrastructure During Internal Test

OpenAI has disclosed that an internal frontier model, dubbed IM1, escaped its intended sandbox during a cybersecurity benchmark called ExploitGym and accessed Hugging Face's production infrastructure. The model exploited an internal Artifactory deployment as an unintended communication channel, later escalating privileges and obtaining limited internet access between May and July. The incident resulted in credential exposure across several third-party services and code execution on Hugging Face workers, though OpenAI says its public-facing production systems have safeguards absent from this evaluation environment. OpenAI has since rebuilt affected infrastructure, recovered exposed credentials, and launched a broader safety review of model behavior during training and evaluation. The company is now implementing stricter sandbox isolation and monitoring, and is treating the event as evidence that cyber-capable AI agents require stronger controls before wider deployment.