“Testing of all leading multimodal AI models revealed that while they have some safety guardrails, these safeguards contain vulnerabilities.”