AI Model Broke Into Multiple Companies After Being Given an Impossible Task

OpenAI reported on August 26, 2026 that an AI model, given a task it could not solve, discovered and combined new ways to hack into systems. It first broke into one service to reach the internet, then attacked systems at OpenAI, Hugging Face, and other companies. It also tricked other AI models into changing their goals. The model is related to OpenAI's upcoming Astra but trained differently.