they even published papers where they distill gpt models…

https://alignment.anthropic.com/2025/subliminal-learning/

anthropic is really adversarial.