MMPIBench finds an audio-agent blind spot

MMPIBench ran 720 tests across six agent frameworks, five models, six visual carriers, and four attacker goals; visual attacks completed in about 1% of runs, but audio attacks completed in 49% of cells when the signal reached a model, according to the benchmark paper. The study says model choice mattered more than framework, with one model never attempting an attack and two attempting in 23.6% of runs. Builders adding voice or image input should threat-model each perceptual channel separately, not inherit visual safeguards by assumption; agent-evaluation governance is the relevant operating precedent.