safety
A Misheard Command Could Slip Past the Safety Check
Simulated speech-recognition errors weakened refusals and allowed unsafe plans from embodied AI systems.
Summary
Simulated speech-recognition errors weakened refusals and allowed unsafe plans from embodied AI systems.
The study combines simulated automatic-speech-recognition errors with SafeAgentBench and POEX to test whether corrupted user input changes embodied-agent behavior. The authors report that some errors preserve enough structure to create harmful ambiguity, while others weaken refusal behavior and permit unsafe plans. Automatic correction reduced risk in some cases but not consistently, so the result identifies an input-channel safety problem rather than a universal correction strategy.
Why it matters
Simulated speech-recognition errors weakened refusals and allowed unsafe plans from embodied AI systems.
Limits and context
- Automatic correction reduced risk in some cases but not consistently, so the result identifies an input-channel safety problem rather than a universal correction strategy.
Key claims
Simulated speech-recognition errors weakened refusals and allowed unsafe plans from embodied AI systems.
Qualification: Automatic correction reduced risk in some cases but not consistently, so the result identifies an input-channel safety problem rather than a universal correction strategy.
Evidence: source-2026-08-31-003
Sources
- arXiv preprint 2608.28518arXiv · primary research
Corrections
No corrections have been recorded for this story.