Prompt-Induced Bias in LLM Security Recommendations: A Quantitative Behavioral Study
Liam Hartley, Emily Zhao, Sophie McMillan · Computer Science Bulletin · 2026
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Computer Science Bulletin · OpenAlex · article · 2026
Abstract
This study explores how prompt phrasing influences the securityrecommendations produced by large language models. Using 4,200controlled prompt variations,we analyze how prescriptive wording, threat emphasis, risk framing, and confidence cues shape LLM suggestions related to input validation, authentication, and cryptographic configurations. Results reveal substantial prompt-induced bias: risk-framing prompts increase conservative recommendations by 31%, while permissive wording reduces strict security advice by 23%. Moreover, confidence-leading prompts cause over-assertive but incorrect recommendations in 12% of cases.
A mitigation framework based on neutralization templates reduces bias variance by 64%. These findings highlight the importance of prompt neutrality for dependable security decision support.
Open the resolved locator · 10.71465/csb222
Metadata licence: OpenAlex metadata (CC0)