Keeping AI models from drifting into malicious behavior.