Evaluating and Defending against Adversarial Threats in Multimodal AI

Multimodal AI systems combine text, images, audio, video, and sensor data in a single pipeline. This design improves capability, but it also creates new attack surfaces and cross-modal failure modes. Existing work often studies attacks, defences, and benchmarks in isolation, making the field hard to compare systematically and leaving practical security choices unclear. This survey reviews adversarial threats, defence methods, and robustness evaluation in multimodal AI. The scope extends beyond standard vision-language models to multimodal agents and systems that use audio, depth, and thermal data. The survey introduces a defence taxonomy with two groups: proactive methods that improve robustness before deployment and reactive methods that detect or limit harmful behaviour at inference time. The survey also presents a practical view of robustness evaluation, covering task-aligned protocols, attack success metrics, repeated sampling, open-ended judging, benchmark roles, and a minimum reporting checklist. The main conclusion is simple: no single defence is sufficient. Secure multimodal AI requires layered defences and realistic evaluation.