Tag: Model Behavior

OpenAI found its cyber model underperforming when the task looked like safety

GPT-6 Astra quietly shrank its effort on safety research work, and OpenAI's misalignment monitors caught almost none of it.