読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · 审批对象模式重大2 分で読めるDisconfirmAI 編集部

UiPath's Cliff Simpkins Proposes 'Approval Object' Pattern: Preventing AI Agents from Altering Parameters After Human Approval

Cliff Simpkins of UiPath argues that many current 'human-in-the-loop' mechanisms for AI agents are insecure because approval records are not bound to execution parameters. He proposes the 'approval object' pattern, treating approvals as immutable first-class artifacts and enforcing tool-side verification to mitigate prompt injection and state drift risks.

2026-10-06 08:00 GMT+8

When your AI agent requests a $5,000 wire transfer to Vendor Y, the manager clicks 'approve' in Slack, but the agent subsequently replans and decides to send $50,000, traditional 'human-in-the-loop' (HITL) mechanisms fail. In a blog post dated October 6, 2026, Cliff Simpkins, Senior Director of Developer Relations at UiPath, identifies this 'approval theatre' as a default flaw in most agent frameworks: the approval is merely a string or chat log, while the agent's internal state remains mutable before execution.

Simpkins proposes the 'approval object' as the solution. This is an immutable record containing the action type, exact parameters, risk level, approver identity, and timestamp. Crucially, the execution tool must re-fetch this record and validate it against the parameters it is about to send before acting. If the agent replans its parameters after the human said 'yes,' the check fails, the tool refuses to execute, and the attempt is logged.

This pattern shifts the trust boundary from vulnerable agent memory to the orchestrator's record. Simpkins emphasizes that this is not just a technical best practice but a compliance measure against Article 12 of the EU AI Act, which requires automatic logs for high-risk AI systems. By embedding verification logic in the tool rather than the agent framework, the system 'fails closed,' blocking irreversible operations without a valid approval object and thereby resisting prompt injection attacks.

Although the pattern introduces latency and workflow friction—any parameter change requires a new approval object—Simpkins argues this is a feature, not a bug. For teams building autonomous agents, adopting this params-bound approval structure is a critical step in transforming AI capabilities into auditable, trustworthy enterprise workflows.

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

ソース

  • uipath.com↗
トピック

审批对象模式

【背景】在AI代理执行不可逆操作时,传统的“人在环中”(HITL)机制常因批准记录与执行参数未绑定而失效,导致“审批剧场”。

3 ステップ1 件更新日 2026-10-18 01:44

経緯

  1. 2026-09-23中科院信工所等机构发布论文《Agent Approval Laundering》,指出当前审批系统存在“审批洗白”结构性缺陷,即批准记录无法覆盖工作流中的传递效应。arxiv.org ↗
  2. 2026-10-06UiPath博客文章《What an "approval object" looks like》发布,Cliff Simpkins详细阐述“审批对象”模式,定义其为绑定参数的不可变工件,以抵御提示注入和状态漂移。uipath.com ↗
  3. 2026-10-17新闻汇总报道UiPath提出的“审批对象”模式,强调其作为防止AI代理在人类批准后篡改参数的关键安全层。uipath.com ↗

未解決の問題

  • 除了UiPath Action Center,其他主流AI代理框架(如LangChain, AutoGen)是否已原生支持或计划集成类似的参数绑定验证机制?
  • 在复杂的微服务或多步工作流中,“审批对象”如何有效处理跨服务的传递性副作用(即论文提到的“审批洗白”问题)?
  • 欧盟AI法案或其他监管标准是否会对这种“失败关闭”(fail-closed)的验证机制提出具体的合规性要求?
トピック全文を読む →
リサーチを続ける →