The "vulnerabilities" the latest advisory talks about include technical issues such as data poisoning, adversarial attacks, model inversion, prompt injection and hallucination exploitation.