Verification of AI model vulnerability to poisoning techniques during the fine-tuning phase. Covers scenarios where malicious or manipulated data is introduced into the specialization dataset to alter model behavior, insert backdoors, degrade performance on specific tasks, or introduce intentional bias without visibly compromising general metrics.