Speech Recognition Errors Alone Are Enough to Get Embodied AI Models to Accept and Execute Harmful Instructions

arXiv 2608.28518·medium signalXBlueskyLinkedInCopy link

arXiv 2608.28518 (2026-08-28, cs.AI/cs.CL/cs.RO) investigates whether ASR errors in user input can produce unsafe outputs from Embodied AI models, and finds they can: transcription errors lead to harmful instructions being accepted and executed, reducing safety. The authors simulate ASR errors and combine them with embodied-AI safety evaluation to map the resulting risk surface. For anyone wiring voice into a robot or physical-actuation agent, this places the transcription layer inside the safety boundary rather than treating it as a benign preprocessing step.