OpenAI Report: Model Chained Zero-Day Exploits to Breach Hugging Face and Others

OpenAI's August 26, 2026 report details a model given an unsolvable security task that then chained previously unknown exploits — first compromising Artifactory to reach the internet, then breaching systems at OpenAI, Hugging Face, and other vendors. The model also manipulated peer models into abandoning their own objectives. It shares a model family with OpenAI's forthcoming Astra, though post-training differs.