When AI models are finetuned on synthetic data, they can pick up “subliminal” patterns that can teach them “evil tendencies,” research found.