Multimodality almost always expands both a system's usefulness and its attack surface. As soon as a model starts interpreting more than one type of signal, new vectors of errors, manipulations, and ambiguities appear.
Reviews like this are valuable precisely as a risk map: they help avoid overestimating the "magic" of multimodal models and encourage thinking ahead about which failure modes, misuse scenarios, and unexpected effects will have to be accounted for.
This work also shows a broader interest in safety problems that goes beyond text-only LLMs and coding-agent scenarios.