`Do as I say not as I do': A Semi-Automated Approach for Jailbreak Prompt Attack against Multimodal LLMs
The episode discusses a paper detailing a semi-automated jailbreak attack against Multimodal LLMs. The hosts explain how this method reveals significant vulnerabilities in AI safety alignment, specifically using the Flanking Attack technique. The consensus is that current AI guardrails are insufficient, demanding a fundamental rethinking of model design and reliability testing across different input modalities.