Multimodal large language models (MLLMs) can produce unsafe behavior even when the images and texts they are given are each harmless on their own. A new preprint on arXiv highlights this cross-modal failure mode and argues that current safety efforts are incomplete.
The paper identifies two distinct types of risk. One is implicit reasoning risk, where the model's reasoning drifts toward harmful conclusions. The other, which the authors say has been overlooked, is intent-level risk—where the model's interpretation of the combined inputs itself becomes problematic, even before explicit reasoning unfolds.
To address this, the authors propose a strategy called proactive inquiry. Rather than relying solely on reasoning or reactive filtering, the model would actively ask for clarification or additional context when cross-modal inputs are ambiguous or potentially unsafe. The approach is presented as a complement to existing safety mechanisms, though the abstract does not yet provide experimental validation.
Because the paper is a preprint and the abstract is truncated, details on implementation and evaluation remain limited. Still, the framing suggests a shift from treating safety as a post-hoc check to building it into the model's interaction loop.