“In safety tests, Anthropic's JLens tool revealed a model trained to misbehave silently ran concepts like "fraud" and "secretly" on ordinary prompts, which were not present in its output.”