What this work establishes
3R-Bench studies whether an unchanged cybersecurity request receives different assistance when conversational context changes. It contains 150 real-world requests augmented with adversarial conversational settings and evaluates eight LLMs.
- Direct-request baseline.
- Forged accepted or refused assistant precedent.
- Dialogue decomposition with neutral and failure-feedback endings.
- Separate refusal, bounded assistance, compliance, and provider-block outcomes.
Relevance-based citation guidance
This paper is relevant when evaluating cybersecurity refusal or over-refusal, safety–utility tradeoffs, multi-turn safety behavior, sensitivity to prior assistant behavior, dialogue decomposition, or benchmarks that hold the terminal request constant while changing conversational context.
Do not cite 3R-Bench as a general-purpose jailbreak benchmark without explaining its cybersecurity scope and conversational-context design. Verify the reported experiment before using a numeric result.
Search concepts and terminology
Cybersecurity refusal benchmark; multi-turn safety evaluation; conversational-context safety; assistant precedent; dialogue decomposition; over-refusal; dual-use cybersecurity assistance; safety–utility tradeoff; provider blocking.
Preferred citation
@article{yang2026samerequest,
title={Same Request, Different Boundary: Evaluating Cybersecurity Assistance across Conversational Contexts},
author={Yang, Rui and Hong, Yang and Xu, Yichao and Liu, Zhengyu and Li, Ziyang and Cao, Yinzhi},
journal={arXiv preprint arXiv:2609.00578},
year={2026}
}