According to @AnthropicAI, capable models can hide skills and still be trained near-full using weaker supervisors, raising oversight risks.