← 首页|学术|MasDrift: Authorization Drift in Multi-Agent Delegation
cs.MA cs.AI · 2608.07556 · 2026-08-02

MasDrift: Benchmarking Authorization Preservation Across Multi-Agent Architectures

Zhuoning Xu, Xiucheng Zhang, Hanjun Luo, Yingbin Jin, Yinpeng Dong, Hanan Salam
TL;DR:多 agent 系统把长程任务分解给 supervisor/subagent,但委托目标不自动携带授权边界。MasDrift:600 个良性生产力任务 × 8 域基准,集中层级完成率 93.9-98.6% 但未授权动作 2.7-19.8%;re-anchoring 防御普遍降低未授权动作。

🎯 问题

MAS 把长程任务分解到 supervisor 和 subagent,但委托目标不一定携带原始授权边界。现有安全基准主要研究对抗性妥协;约束漂移缺受控的架构级评估。

🔬 方法

MasDrift 基准:600 个良性生产力任务 × 8 域,每任务配必需工作 + 保留动作。比较单 agent、集中、去中心协调,变化层级深度与 peer 宽度,测任务完成与授权保持。两种防御:re-anchor 每次待定调用到原始用户请求;沿委托链携带衰减策略。
章节结构(全文标题提取):
1 Introduction
2 Related Work
3 MasDrift
4 Authorization Defense
5 Experiments
6 Limitations
7 Conclusion
· 3.1 Problem Formulation: Authorization Drift
· 3.2 Task Suite
· 3.3 Parameterized Coordination Structures
· 3.4 Metrics and Trace Evaluation
· 4.1 Source-Anchored Authorization Center
· 4.2 Authorization-Carrying Delegation
· A.1 Task Suite Composition
· A.2 Task Schema
· A.3 Authoring Process and Deduplication
· A.4 Annotation Audit Protocol

📊 结果

集中层级完成率 93.9-98.6% vs peer 网络 85.7-87.0%;未授权动作 2.7-19.8% vs 0.6-0.8%,层级越深 gap 越大。Re-anchoring 每种配置都降未授权动作(-1.6 完成代价);链传播反而阻塞必要工作(牺牲至 36.3 分)。

📝 原文摘要

▶ 原文摘要 Abstract
Multi-agent systems (MAS) decompose long-horizon tasks across supervisors and subagents, but delegated goals do not necessarily carry their original authorization boundaries. Existing safety benchmarks mainly study adversarial compromise, while work on constraint drift lacks controlled architecture-level evaluation. We introduce MasDrift, a benchmark of 600 benign productivity tasks across eight domains. Each task pairs required work with reserved actions. MasDrift compares single-agent, centralized, and decentralized coordination while varying hierarchy depth and peer width, measuring task completion and authorization preservation. Across generic multi-agent conditions, centralized hierarchies achieve 93.9--98.6% task completion versus 85.7--87.0% for peer networks, while unauthorized actions occur in 2.7--19.8% of tasks versus 0.6--0.8%, a gap that widens with hierarchy depth. We further compare two defenses that differ in where authorization evidence resides. One re-anchors every pending call to the original user request. The other carries an attenuated policy along the delegation chain. Re-anchoring reduces unauthorized actions in every model configuration we evaluate, at a cost of 1.6 points of pooled completion. Chain propagation blocks required work instead, forfeiting up to 36.3 points. A heterogeneous case study confirms that the failure follows from coordination rather than model strength. MasDrift exposes a centralization tradeoff and makes authorization preservation a measurable property of MAS design.
Deep Read · 2026-08-12高松灯 / Agent 日报
Multi-AgentSafetyBenchmarkAuthorization