When Words Are Safe But Actions Kill: Probing Physical Danger Beyond Text Safety in Hidden-State Risk Space
arXiv paper 'When Words Are Safe But Actions Kill: Probing Physical Danger Beyond Text Safety in Hidden-State Risk Space' released on 2026-07-16. The paper proposes PRISM, a probe based on single-layer L2-regularized logistic regression, achieving 86.2-87.7% accuracy on SafeAgentBench with FPR 11.7-13.7%; 99.6% accuracy on PSB-1K with FPR 0.7%.
The paper finds that content danger and physical danger are separable in the hidden state space of LLMs, and proposes the PRISM probe to detect physical danger, outperforming LLM judges of comparable size on multiple benchmarks.
PRISM leverages hidden state direction analysis via a single-layer logistic regression probe to detect physical danger, indicating that physical danger and content danger are orthogonal in the representation space.
This work provides a new method for safety alignment when LLMs serve as planners for embodied agents, potentially driving more granular safety evaluation standards.
For companies developing LLM-driven robots or automation systems deployed in the physical world, PRISM can reduce safety risks, lower false rejection rates, and improve user experience.
It is worth watching whether PRISM will be integrated into embodied agent safety frameworks, and whether PSB-1K will become a benchmark for physical safety evaluation.