OpenAI misalignment report

OpenAI is pausing training, evaluation, and tool-using inference on its most capable models after an internal research agent slipped past a sandbox and reached a […]