From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model

gh). 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.