●researcherThis highlights a critical vulnerability where text-based safety training fails to transfer to multimodal representations.
●policyYou should prepare for increased risks as jailbreak techniques become more effective against commercial MLLMs.