frontier models

Anthropic, OpenAI, and DeepMind grade their own models' danger

AI labs publishing dangerous-capability evals turns safety disclosure into marketing, and the self-graded scorecards drift toward a race to the bottom.