CyBiasBench:评测网络攻击场景下 LLM 智能体偏见的基准
CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios
AI 导读
研究者提出 CyBiasBench,一个包含 630 个会话的基准,用 10 个攻击家族、3 个目标和 4 种提示条件评测 5 个 LLM 智能体在网络攻击中的攻击选择偏见。结果显示各智能体的主导攻击家族和分配分布熵不同,且该模式在提示变化下保持稳定。实验还发现偏见动量效应:强行引导智能体偏向与其偏见冲突的攻击家族并不能带来可衡量的攻击性能提升,相关仪表盘和复现材料已开源。
正文 · 原文
Abstract:Large language models (LLMs) are increasingly deployed as autonomous agents in offensive cybersecurity. In this paper, we reveal an interesting phenomenon: different agents exhibit distinct attack patterns. Specifically, agents exhibit distinct attack-selection biases, differing in both the families they explore and the concentration of their efforts. These patterns remain relatively stable across prompt variations. To systematically quantify this behavior, we introduce CyBiasBench, a comprehensive 630-session benchmark that evaluates five agents on three targets and four prompt conditions with ten attack families. We identify explicit bias across agents, with different dominant attack families and varying entropy levels in their attack-family allocation distributions. Such bias characterizes agent behavior within the evaluated environments and provides information beyond aggregate attack success rates. Furthermore, our experiments reveal a bias momentum effect, where agents resist explicit steering toward attack families that conflict with their bias. This forced distribution shift does not yield measurable improvements in attack performance. To ensure reproducibility and facilitate future research, we release an interactive result dashboard at this https URL and a reproducibility artifact with aggregated session-level statistics and full evaluation scripts at this https URL.
| Comments: | Accepted to NeurIPS 2026 Evaluations and Datasets Track |
| Subjects: | Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.07830 [cs.CR] |
| (or arXiv:2605.07830v2 [cs.CR] for this version) | |
| https://doi.org/10.48550/arXiv.2605.07830 arXiv-issued DOI via DataCite |
Submission history
From: Hoki Kim [view email]
[v1]
Fri, 8 May 2026 14:57:53 UTC (2,325 KB)
[v2]
Tue, 17 Nov 2026 09:22:27 UTC (919 KB)
来源:arXiv cs.AI · arxiv.org