When Words Are Safe But Actions Kill: Probing Physical Danger Beyond Text Safety in Hidden-State Risk Space

Researchers have identified a critical disconnect between the safety of text generated by large language models and the potential physical danger that can arise when those instructions are executed in the real world. A recent study reveals that even linguistically benign text can become hazardous when grounded in physical actions, highlighting a distinct safety concern that goes beyond traditional text-level content danger. The investigation utilized hidden-state direction analysis and random-split null tests to probe this phenomenon, demonstrating that physically grounded danger is a separate safety problem. This finding has significant implications for the development and deployment of embodied agents that rely on large language models for planning. The study's results underscore the need for a more nuanced understanding of safety risks in AI systems, particularly as they are integrated into physical environments. This matters to practitioners because it emphasizes the importance of considering physical safety risks when designing and implementing AI-powered systems¹.