Definition
The extra ways an AI system can be attacked because it accepts more than just text — such as images, audio, or video — each requiring its own processing code that can contain its own bugs. Attackers can exploit how these extra formats are decoded to crash the system, exhaust resources, or slip in hidden malicious content.
Why it matters
As AI products race to add voice and image features to compete, each new input type quietly expands what a security team must defend, often faster than defenses can catch up — as shown by repeated audio/video-decoding flaws in a widely deployed inference engine.