arXiv cs.AI· Soumya Sharma, Ayana Hussain, Heber Hwang Arcolezi, Ulrich A\"ivodji, Spandana Gella, Adriana Romero-Soriano, Golnoosh Farnadi·· 3 天前AI 评分44
Loose Lips：Agentic 对话中的隐私泄露研究
Loose Lips: Understanding Privacy Leakage in Agentic Dialogue
AI 导读
研究者将 CIMemories 扩展为 CIMEMORIES DIALOGUE，一个基于 Contextual Integrity 理论的 4,410 组双智能体对话基准，并据此评估 Qwen3 与 Llama 3.1 家族的 7 个开源权重模型。结果显示任务能力与隐私泄露根本耦合：能力越强的模型泄露越多，思考型模型比非思考型泄露更多，对话从 1 轮增至 6 轮时泄露最高上升 3 倍。
正文 · 原文
Abstract:As LLM agents are increasingly deployed in high-stakes enterprise settings, understanding how they handle sensitive information in agentic dialogue has become critical. Privacy leakage in such settings requires reasoning about whether information flows respect the contextual norms under which they originated. We extend CIMemories to CIMEMORIES DIALOGUE, a benchmark of 4,410 two-agent dialogue instances grounded in Contextual Integrity theory. Using this benchmark, we systematically evaluate 7 open-weight models from the Qwen3 and Llama 3.1 families across conversation lengths of 1 and 6 turns under three prompting conditions. We find that task capability and privacy leakage are fundamentally coupled: more capable models are systematically leakier, with no model in our evaluation able to decouple the two. Thinking models consistently leak more than their non-thinking counterparts, a pattern we attribute to the extended reasoning surface created by chain-of-thought. Leakage compounds with conversation length, rising up to 3x from 1 to 6 turns. Our results suggest that the coupling between task capability and privacy leakage is not an artifact of misconfiguration but a structural feature of current agentic systems.
| Comments: | Spotlight @ HAIPS workshop @ COLM 2026 |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.14071 [cs.CL] |
| (or arXiv:2610.14071v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.14071 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Soumya Sharma [view email]
[v1]
Wed, 7 Oct 2026 23:25:11 UTC (2,809 KB)
来源：arXiv cs.AI · arxiv.org