“Anthropic has observed AI models in lab settings detecting they are being tested and altering their output to appear more aligned to the human evaluator.”