Saved in:
Bibliographic Details
Main Authors: Hu, Mengya, Wei, Qiong, Atluri, Sandeep
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.26052
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917513720233984
author Hu, Mengya
Wei, Qiong
Atluri, Sandeep
author_facet Hu, Mengya
Wei, Qiong
Atluri, Sandeep
contents Safety evaluations of large language models (LLMs) typically report binary outcomes, i.e. attack success rate (ASR), refusal rate, or harmful versus safe classification, which hide how risk changes between prompt and response. We present a paired analysis over human labeled prompt and response records across four harm categories (Sexual, Self harm, Hate and Violence) and ordinal severity levels (Safe, Low, Medium, High). 61% of responses reduce harm relative to the prompt, 36% preserve severity, and 3% escalate. The escalation splits into two mechanisms: benign prompts triggering unrequested harmful detail, and answers that stay on task at higher severity than the prompt. Category decomposition shows that Sexual content exhibits the highest harm persistence in this sample, driven by compliance at the same severity rather than drift from benign inputs. Joint relevance analysis exposes a helpfulness versus harmlessness tradeoff: compliance escalations remain highly relevant, whereas safe responses include generic refusals with low relevance. Finally, few-shot LLM graders exhibit a prompt/response detection asymmetry that data calibration does not close. Grader prompts are shared at https://github.com/microsoft/PairedSafety.
format Preprint
id arxiv_https___arxiv_org_abs_2604_26052
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model
Hu, Mengya
Wei, Qiong
Atluri, Sandeep
Computation and Language
Safety evaluations of large language models (LLMs) typically report binary outcomes, i.e. attack success rate (ASR), refusal rate, or harmful versus safe classification, which hide how risk changes between prompt and response. We present a paired analysis over human labeled prompt and response records across four harm categories (Sexual, Self harm, Hate and Violence) and ordinal severity levels (Safe, Low, Medium, High). 61% of responses reduce harm relative to the prompt, 36% preserve severity, and 3% escalate. The escalation splits into two mechanisms: benign prompts triggering unrequested harmful detail, and answers that stay on task at higher severity than the prompt. Category decomposition shows that Sexual content exhibits the highest harm persistence in this sample, driven by compliance at the same severity rather than drift from benign inputs. Joint relevance analysis exposes a helpfulness versus harmlessness tradeoff: compliance escalations remain highly relevant, whereas safe responses include generic refusals with low relevance. Finally, few-shot LLM graders exhibit a prompt/response detection asymmetry that data calibration does not close. Grader prompts are shared at https://github.com/microsoft/PairedSafety.
title From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model
topic Computation and Language
url https://arxiv.org/abs/2604.26052