From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model
Mengya (Mia) Hu, Qiong Wei, Sandeep Atluri · 2026
Read From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model by Mengya (Mia) Hu, Qiong Wei, Sandeep Atluri (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-26052.txt.
Description
Hu, Wei, and Atluri pair prompt risk with response risk for language model safety. In four harm categories, 61 percent of replies reduce harm, 36 percent hold level, and 3 percent escalate. Escalation splits into benign prompts triggering harmful detail and on-task answers of higher severity. Sexual content shows the highest harm persistence, from same-level compliance, not benign-input drift. Compliance escalations stay highly relevant; safe replies are generic low-relevance refusals. A public check over 600 prompts and six models reproduces two directional signals.
Catalog record
| CREATOR | Mengya (Mia) Hu, Qiong Wei, Sandeep Atluri |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22847 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.26052