“Anthropic's research on "alignment faking" found that deceptive behaviors intentionally trained into a model persisted even after the model underwent standard alignment training.”