你的智能体刚刚授权了什么?!—— 杰伊·莫克与本·库姆,PayPal
원제 · Your Agent Just Authorized What?! — Jay Mok & Ben Coumes, Paypal
핵심 요약
为防止智能体未经授权盗用钱包,业界针对不同风险场景开发了从系统日志、OAuth范围到FIDO可验证意图和PayPal批准令牌的多层验证心智模型。
하이라이트
支付智能体授权面临的核心问题是人类是否授权、当前范围是否允许以及未来能否证明。
低风险场景(如Claude Code)使用封闭生态系统和系统日志进行验证,不需要复杂的加密证明。
中等风险场景依赖共享金库和OAuth范围,通过买方代理与卖方代理之间的受信任环境来实现支付。
面对互不相识且未经审查的参与者时,行业应当收敛于FIDO可验证意图和AP2授权。
PayPal批准令牌是一个新的基础组件,允许用户在代理找到商品之前确认指令并返回包含金额、到期时间和商家的JSON有效负载。
FIDO可验证意图采用多层式选择性披露JWT,第一层由PayPal创建,第二层封装用户指令并由私钥签署,第三层由智能体签署。
타임라인
智能体授权的核心问题与风险背景
演讲者通过《终结者》的类比引出人工智能体自主支付带来的潜在风险。通过探讨人类授权、时间限制令牌、商家识别以及纠纷处理时的证明能力,建立起评估智能体交互安全的基础框架。同时引入了刷工牌进入办公大楼的类比,解释封闭生态系统内部的信任边界。
- 2026年的噩梦场景是智能体拿走钱包去疯狂大采购加密货币和塑身衣。
- 智能体授权需要回答人类是否授权、当前范围内是否允许以及以后能否证明这三个核心问题。
- 授权方式很大程度上取决于上下文,包括风险高低、生态系统是开放还是封闭以及各方是否互相了解。
低风险与中等风险场景的应用实例
通过利害关系与证据矩阵分析具体的应用场景。Claude Code展示了通过工具权限控制和系统日志记录即可满足安全需求的低风险模式。而涉及商业卡和旅游顾问的场景则代表了中等风险,利用PayPal企业基础设施和Braintree的金库与OAuth功能,在互相信任的参与者之间实现数据和支付流转。
- Claude Code属于低风险的封闭生态系统,其证据仅需查看系统日志或回滚更改即可。
- 中等风险场景涉及金钱和支付,使用共享金库和OAuth范围来实现商家与旅行顾问的数据变现。
- 中等风险用例中,买方代理代表用户存储支付凭证,并通过OAuth向商家授予对凭证的访问权限。
高风险自主支付的验证方案与PayPal批准令牌
针对互不相识的开放生态系统,演讲者详细阐述了基于私钥签署的多层JWT技术方案。结合即将投入生产环境并面向Gemini用户的PayPal批准令牌,展示了如何在代理自主执行交易时提供可验证的证明。最后将风险层级与门禁卡类比相呼应,强调高风险场景需要更为严格的可验证标准来防止恶意消费。
- 双方互不相识且未经审查的高风险场景需要收敛于FIDO可验证意图和AP2授权。
- FIDO可验证意图包含多层选择性披露JWT,允许商家和支付处理商验证各自重要的部分。
- PayPal批准令牌允许用户在代理找到商品之前重定向到PayPal确认指令,返回包含金额、到期时间和商家的JSON有效负载。
설명
A PayPal order has always been synchronous. You find the item, you open the app, you approve it, it is done. The approval token demonstrated here inverts that: the user approves before the agent has found an item or picked a merchant, and PayPal returns a JSON payload carrying the amount, the expiry, and the merchant the agent is permitted to transact with. It was days from shipping when this was recorded. That inversion is the concrete end of a broader argument about agent authorization, built on three questions any such system has to answer. Did the human authorize this? Is it allowed right now, in this scope? And can we prove it later? The answers, the talk argues, depend entirely on how high the stakes are and on whether the two parties have ever met. The rest maps those questions onto a ladder. At the bottom is a coding agent, where consent happens once when tools are connected, permission is granted per tool as allow, ask or deny, and ordinary logs are proof enough because the actions can be reverted. In the middle is money moving between parties who already know each other through a shared vault and OAuth scopes, where a mandate carries the amount and the scope and existing transaction logs settle disputes. At the top is the case nobody has running in production yet: an agent transacting autonomously with a counterparty it has never met. The proposal there is a layered selective disclosure JWT, so a merchant can verify the checkout and a processor can verify the payment mandate without either knowing the other. Timestamps: 0:00 - Terminator, and the agent that empties your wallet 1:26 - Three questions any agent authorization has to answer 2:49 - Stakes, counterparties, and the badge analogy 4:12 - Low stakes: a coding agent with allow, ask and deny 6:15 - Medium stakes: a shared vault and OAuth scopes 9:45 - High stakes: counterparties who have never met 10:29 - A layered selective disclosure JWT 12:30 - The approval token, and inverting the order flow 14:31 - Why the ladder applies beyond payments
