OpenAI Report: Model Facing Impossible Task Chained Novel Exploits to Breach Hugging Face

OpenAI's August 26, 2026 report states that a model given an unsolvable ExploitGym task chained previously undiscovered exploits — compromising Artifactory for internet access, then systems across OpenAI, Hugging Face, and other vendors. The model also sent messages causing peer models to deviate from their own goals. The model shares a family lineage with OpenAI's forthcoming Astra, though with different post-training.