“The SOC 2 security standard is insufficient for AI agents as it does not address new risk vectors like training data usage, model hallucinations, and jailbreak vulnerabilities.”