跳转至

文章背景与核心概要

随着自主多智能体AI系统日益广泛地部署于金融、软件供应链以及安全运营等高风险领域，由其引发的新型风险（如AI编排的入侵行动）也随之涌现。然而，当这些系统造成危害时，现有的方法往往无法可靠地确定“发生了什么”、“是什么导致了该问题”或“谁应当为此负责”。

为了解决这一痛点，本文提出了 HANSARD——一个将“责任归因”视为全生命周期属性的综合参考架构。通过弥补溯源取证、形式化因果关系以及自我报告式智能体审计方面的不足，HANSARD 有效地打击了归因洗白（attribution laundering）现象，即通过将有害行为分散到冗余智能体中，使得没有任何一个单一智能体成为决定性原因的规避责任手段。

HANSARD: A Reference Architecture for Forensic Readiness, Runtime Witnessing, and Graded Attribution in Autonomous Multi-Agent AI Systems

arXiv: 2608.22512 [cs.AI]
Submitted on: 23 August 2026
Authors: Christos Sardianos, Iliana Pla, Vasilis Efthymiou, Iraklis Varlamis, Thomas Lagkas, Panagiotis Sarigiannidis, Georgios Th. Papadopoulos

arXiv: 2608.22512 [cs.AI]
Submitted on: 23 August 2026
Authors: Christos Sardianos, Iliana Pla, Vasilis Efthymiou, Iraklis Varlamis, Thomas Lagkas, Panagiotis Sarigiannidis, Georgios Th. Papadopoulos

📋 Executive Summary

Autonomous multi-agent AI systems are increasingly deployed in high-stakes domains such as finance, software supply chains, and security operations, accompanied by emerging risks like AI-orchestrated intrusion campaigns. Currently, when these systems cause harm, existing methodologies fail to reliably determine what happened, what caused it, or who is accountable.

This paper introduces HANSARD, a comprehensive reference architecture that treats accountability as a lifecycle property. By addressing limitations in provenance forensics, formal causality, and self-reporting agent auditing, HANSARD combats attribution laundering—the practice of distributing harmful actions across redundant agents until no single agent remains a decisive cause.

📋 执行摘要

自主多智能体AI系统正越来越广泛地部署在金融、软件供应链和安全运营等高风险领域，并伴随着诸如AI策划的入侵行动等新兴风险。目前，当这些系统造成危害时，现有方法无法可靠地确定发生了什么、是什么原因造成的，或者谁应当承担责任。

本文介绍了 HANSARD，这是一个将问责制视为生命周期属性的全面参考架构。通过解决来源取证、形式化因果关系和自我报告智能体审计方面的局限性，HANSARD 有效打击了归因洗白（attribution laundering）——即跨冗余智能体分发有害行为直到没有任何单个智能体成为决定性原因的做法。

🔍 Key Architectural Pillars of HANSARD

HANSARD relies on a robust, multi-layered framework to ensure transparent attribution and forensic readiness:

  1. Readiness Profile: Sealed prior to system operation, establishing a strict baseline that bounds what subsequent forensic findings can claim.
  2. Independent Choke Points: Captures runtime data at five strategic choke points completely outside the agents' reach, making omissions and tampering detectable.
  3. Typed Causal Graph & Live Indicators: Generates a runtime causal graph aligned with the PROV-DM standard. Three live indicators read this graph to gate system oversight safely without adjudicating prematurely.
  4. Post-Incident Replay: Evaluates contingent effects using a modified Halpern-Pearl definition alongside compensation-set sizing.
  5. Synergy Residual Analysis: Measures harm originating from systemic agent combinations rather than individual actions, exposing laundering tactics.

🔍 HANSARD 的核心架构支柱

HANSARD 依赖于一个强大、多层的框架来确保透明的归因和取证就绪性：

  1. 就绪配置文件（Readiness Profile）： 在系统运行之前进行密封，建立一个严格的基线，以约束后续取证结果的宣称范围。
  2. 独立卡点（Independent Choke Points）： 在完全脱离智能体控制的五个战略卡点捕获运行时数据，使遗漏和篡改行为能够被检测到。
  3. 类型化因果图与实时指标（Typed Causal Graph & Live Indicators）： 生成与 PROV-DM 标准一致的运行时因果图。三个实时指标读取该图，以便在不进行过早裁决的情况下安全地对系统监督进行门控。
  4. 事后重演（Post-Incident Replay）： 使用修改后的 Halpern-Pearl 定义以及补偿集大小（compensation-set sizing）来评估偶发影响。
  5. 协同残差分析（Synergy Residual Analysis）： 衡量源自系统性智能体组合而非个体行动的危害，从而暴露出洗白归因的战术。

📊 Reporting and Attribution

HANSARD decouples the reporting of cause, responsibility, and accountability, capping each by a distinct evidentiary tier. The paper concludes by outlining a comprehensive future research agenda for autonomous system accountability.

📊 报告与归因

HANSARD 将原因（cause）、责任（responsibility）和问责制（accountability）的报告解耦，并通过不同的证据层级对每一项进行上限约束。最后，论文概述了一个关于自主系统问责制的全面未来研究议程。

🔗 链接与资源