Multimodal agents increasingly act on what they see, not just what they read. That shift creates a safety gap: text-only evaluations may miss attacks delivered through visual interfaces. A new preprint, MMA-SafetyBench, proposes a benchmark aimed at closing that gap.
At its center is Visual Semantic Hijacking, a class of attacks in which deceptive visual content is used to mislead an agent's perception or decision-making. The authors argue that because agents rely on visual interfaces to perceive state and select actions, these inputs form a distinct attack surface.
The abstract is truncated, so the full methodology and results are not yet available from the announcement. The paper is posted on arXiv as 2610.19384v1.