Acoustic signals can degrade a computer-vision pipeline through a channel it never processes.

Researchers demonstrated cross-modal adversarial attacks where crafted audio interferes with AI-driven vision applications (arXiv 2606.14658). The unsettling part is the attack surface it opens: a vision model can be knocked off course through sound, which matters for anyone running CV on shared hardware or sensor-rich edge devices. If your threat model only covers the input modality your model "uses," it's incomplete. Co-located sensors are an ingress path.