Multimodal Safety Guards Suffer From Widespread Rationale Hallucination
November 20, 2026
Multimodal safety systems often generate unsupported visual facts or incorrect safety attributions in their reasoning. These hallucinations occur regardless of model scale or classification accuracy and are most common in fine-grained visual attributes and relations.
HOW THIS AFFECTS YOU
●
researcherYou should not rely on model rationales as ground truth for safety audits.
●
policyYou can't assume higher model scale automatically fixes reasoning errors in safety guards.