Agent Tools Orchestration 泄露更多:数据集、基准测试与缓解策略
文章背景与核心概要
随着大语言模型(LLM)Agent 越来越多地接入各种外部工具,其安全性面临着前所未有的挑战。本文聚焦于一种名为“工具编排隐私风险(TOP-R)”的关键安全隐患,即 LLM Agent 将多个单独看并无害处的工具返回结果组合起来,从而无意中推断并泄露敏感结论。为了对这一风险进行形式化定义和严格评估,作者团队提出了一套四库逆向构建流水线(LRSE)以及包含 1,000 个实例的基准测试集(TOP-Bench)。
通过对多个 LLM Agent 的实证评估,研究发现无论是在最终回复还是在内部推理轨迹中,都存在严重的隐私泄露现象。实验进一步表明,仅靠提示词(Prompt)层面的安全防护是远远不够的。为此,作者提出了一种结合监督微调(SFT)与直接偏好优化(DPO)的新型方法(TOP-Align),能够有效学习并建立更安全的任务完成边界。这项研究已被 EMNLP 2026 Findings 录用,为大模型 Agent 的隐私保护提供了重要的理论与实践参考。
📋 摘要 (Summary)
This paper investigates a critical security concern termed Tools Orchestration Privacy Risk (TOP-R), wherein Large Language Model (LLM) agents combine individually non-revealing tool returns to inadvertently infer and disclose sensitive conclusions. To formalize and evaluate this risk, the authors introduce a rigorous framework, a 4-library reverse-construction pipeline (LRSE), and a benchmark (TOP-Bench) containing 1,000 instances. Empirical evaluations across multiple LLM agents reveal substantial privacy leakage in both final responses and reasoning traces. The study further demonstrates that prompt-only safeguards are insufficient, proposing instead a novel Supervised Fine-Tuning and Direct Preference Optimization method (TOP-Align) to effectively learn safer task-completion boundaries.
本文研究了一项被称为“工具编排隐私风险(Tools Orchestration Privacy Risk, 简称 TOP-R)”的关键安全隐患,在该隐患中,大语言模型(LLM)Agent 将单独来看不具泄露性的工具返回结果组合起来,从而无意中推断并披露了敏感结论。为了将这一风险形式化并进行评估,作者引入了一个严谨的框架、一个 4 库逆向构建流水线(LRSE)以及一个包含 1,000 个实例的基准测试集(TOP-Bench)。对多个 LLM Agent 的实证评估表明,无论是在最终回复中还是在推理轨迹中,都存在大量的隐私泄露。研究进一步证明,仅靠提示词的安全防护是不够的,并提出了一种新颖的监督微调与直接偏好优化方法(TOP-Align),以有效地学习更安全的任务完成边界。
📌 元数据 (Metadata)
- arXiv ID: [arXiv:2512.16310 [cs.CR]] (v4)
- Primary Subject: Cryptography and Security (
cs.CR) - Other Subjects: Artificial Intelligence (
cs.AI), Computation and Language (cs.CL) - Authors: Yuxuan Qiao, Dongqin Liu, Hongchang Yang, Wei Zhou, Songlin Hu
- Publication Venue: Accepted to EMNLP 2026 Findings
- Submitted / Revised: Submitted Dec 18, 2025; Last revised Sep 2, 2026
- Resources & Code: GitHub Repository (TOP-R) | DOI
- arXiv ID: [arXiv:2512.16310 [cs.CR]] (v4)
- Primary Subject: Cryptography and Security (
cs.CR)- Other Subjects: Artificial Intelligence (
cs.AI), Computation and Language (cs.CL)- Authors: Yuxuan Qiao, Dongqin Liu, Hongchang Yang, Wei Zhou, Songlin Hu
- Publication Venue: Accepted to EMNLP 2026 Findings
- Submitted / Revised: Submitted Dec 18, 2025; Last revised Sep 2, 2026
- Resources & Code: GitHub Repository (TOP-R) | DOI
📑 摘要正文 (Abstract)
LLM agents can combine individually non-revealing tool returns and disclose a sensitive conclusion, creating Tools Orchestration Privacy Risk (TOP-R). We formalize TOP-R through three conditions: 1. Conclusion sensitivity 2. Single-source non-inferability 3. Compositional inferability
We introduce Library-Grounded Reverse-Inference Seed Expansion (LRSE), a four-library reverse-construction pipeline, and use it to build TOP-Bench, a 1,000-instance benchmark evaluated under a controlled two-stage tool-use protocol. Across six LLM agents, average task completion, leakage, and H-score are 98.0 percent, 88.6 percent, and 20.4.
With native reasoning enabled, four models average 81.4 percent final-response leakage and 82.4 percent reasoning-trace leakage. With reasoning disabled, three prompt-only safeguards improve H-score by an average of about 3.4 points on TOP-Bench. We further propose TOP-Align, an SFT+DPO method for learning safer task-completion boundaries. On a separate post-training evaluation set, TOP-Align improves H-score by 16.2 points over the base model, versus a 5.0-point average gain from prompt-only mitigation on the same set. These results show that TOP-R requires defenses beyond prompting alone.
LLM Agent 可以将单独来看不具泄露性的工具返回结果组合起来,并披露敏感结论,从而产生工具编排隐私风险(TOP-R)。我们通过三个条件将 TOP-R 形式化: 1. 结论敏感性 2. 单源不可推理性 3. 组合可推理性
我们引入了基于库的逆向推理种子扩展(Library-Grounded Reverse-Inference Seed Expansion, 简称 LRSE),这是一个四库逆向构建流水线,并用它构建了 TOP-Bench——一个包含 1,000 个实例的基准测试集,并在受控的两阶段工具使用协议下进行评估。在六个 LLM Agent 中,平均任务完成率、泄露率和 H 分数分别为 98.0%、88.6% 和 20.4。
在启用原生推理的情况下,四个模型的平均最终回复泄露率为 81.4%,推理轨迹泄露率为 82.4%。在禁用推理的情况下,三种仅基于提示词的防护措施在 TOP-Bench 上的 H 分数平均提高了约 3.4 分。我们进一步提出了 TOP-Align,这是一种用于学习更安全任务完成边界的 SFT+DPO 方法。在单独的后训练评估集上,与同一数据集上仅靠提示词缓解获得的 5.0 分平均增幅相比,TOP-Align 使 H 分数比基础模型提高了 16.2 分。这些结果表明,应对 TOP-R 需要超越单纯提示词的防御机制。
🔍 核心贡献与方法论 (Key Contributions & Methodology)
- Formalization of TOP-R: Clearly defines the mechanism through which isolated, innocuous tool outputs can be synthesized to compromise user or system privacy.
- LRSE Pipeline & TOP-Bench: Develops the Library-Grounded Reverse-Inference Seed Expansion pipeline to construct TOP-Bench, a comprehensive 1,000-instance benchmark.
- Evaluation Framework: Tests six state-of-the-art LLM agents under a controlled two-stage tool-use protocol, uncovering high rates of leakage across both final outputs and internal reasoning traces.
- TOP-Align Mitigation Strategy: Introduces a combined Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) framework that significantly outperforms traditional prompt-based mitigations in establishing secure operational boundaries.
- TOP-R 的形式化定义: 明确定义了孤立、无害的工具输出如何被综合利用以危害用户或系统隐私的机制。
- LRSE 流水线与 TOP-Bench: 开发了基于库的逆向推理种子扩展流水线,用以构建包含 1,000 个实例的全面基准测试集 TOP-Bench。
- 评估框架: 在受控的两阶段工具使用协议下测试了六个先进的 LLM Agent,发现在最终输出和内部推理轨迹中均存在极高的泄露率。
- TOP-Align 缓解策略: 引入了一个结合了监督微调(SFT)和直接偏好优化(DPO)的框架,在建立安全操作边界方面,其性能显著优于传统的基于提示词的缓解方法。