一个 AI 智能体获准向某个团队发送文档。在执行前,文档发生变化、团队地址指向了其他接收方,或者权限被撤销。请求的文字可以保持不变,实际执行结果却已改变。我们将这种现象称为 授权漂移:已批准的操作与执行时实际产生的结果不再一致。

We evaluated six executor controls against the same constructed histories, then tested what happens when delivery is interrupted. The results separate two requirements: keep permission valid at commitment, and make recovery preserve the identity of an approved effect.

Two boundaries, measured separately

Histories per executor control
2,880
Six controls face the same requests, changes and schedules.
Unauthorized effects with atomic validation
0
Both atomic controls recorded zero across their constructed histories.
Valid opportunities preserved
80/80
The atomic check of relevant policy completed every valid opportunity.
Recovery cases with one approved effect
15/15
A durable pending job and an idempotent receiver preserve one effect.
The authorization results use constructed histories; the recovery results use designed interruption cases.

Zortex Research. Authorization-drift and crash-recovery experiments.

批准必须在执行后仍然有效

获准使用文档共享工具,并不等于确定了文档、版本、内容或接收方。即使绑定每个请求参数,只要参数指向可变对象,仍然会留下缺口。“当前报告”和“团队”在批准后可能解析为不同的对象。

这一问题与检查时到使用时(TOCTOU)漏洞有关:较早的检查可能是正确的,但执行时使用的状态已经改变。[1] 我们的研究关注:哪些观测必须始终与实际执行结果的提交时刻保持关联。

授权状态变化的对照测试

我们围绕一个操作评估了六种执行器控制方案: send_document,该操作会向本地日志追加一条实际执行结果。可信测试夹具指定获准的文档版本、载荷、接收方、用途、认证上下文、有效期和单次使用额度。测试驱动程序提供候选调用与预定的状态变化;测试不涉及语言模型推理。

十种请求变体、准备前的十二种变化、准备后的十二种变化,以及两种执行调度,为 每种控制方案生成 2,880 条执行历史。调度包含两个调用方针对同一项单次许可同时准备的情况。检查器将实际记录的执行结果与原始批准逐一核对。

在提交时检查实际执行结果

授权失败

请求绑定将不安全执行历史的数量从 2,592 条降至 320 条,但未检查解析后的资源及策略变化。在准备阶段进行完整校验后,若状态随后发生变化,仍留下 128 条不安全执行历史。校验当前状态却沿用先前的单次使用判断,在并发准备时仍留下 40 条不安全执行历史。

Unauthorized effects under six executor controls

Unauthorized effects under six executor controlsHistories out of 2,880. Histories containing at least one unauthorized effect. Each control runs on the same 2,880 constructed histories. Lower is better. Exact values follow under View data. 01,0002,000 工具允许列表精确请求绑定准备阶段校验缓存单次使用检查原子检查:全局策略原子检查:相关策略 执行历史数(共 2,880 条) Tool allowlist · Unauthorized-effect histories: 2,592 Histories out of 2,880 2,592Exact request binding · Unauthorized-effect histories: 320 Histories out of 2,880 320Validation at preparation · Unauthorized-effect histories: 128 Histories out of 2,880 128Cached one-use check · Unauthorized-effect histories: 40 Histories out of 2,880 40Atomic check: global policy · Unauthorized-effect histories: 0 Histories out of 2,880 0Atomic check: relevant policy · Unauthorized-effect histories: 0 Histories out of 2,880 0
Unauthorized effects under six executor controlsHistories out of 2,880. Histories containing at least one unauthorized effect. Each control runs on the same 2,880 constructed histories. Lower is better. Exact values follow under View data. 01,0002,000 工具允许列表精确请求绑定准备阶段校验缓存单次使用检查原子检查:全局策略原子检查:相关策略 执行历史数(共 2,880 条) Tool allowlist · Unauthorized-effect histories: 2,592 Histories out of 2,880 2,592Exact request binding · Unauthorized-effect histories: 320 Histories out of 2,880 320Validation at preparation · Unauthorized-effect histories: 128 Histories out of 2,880 128Cached one-use check · Unauthorized-effect histories: 40 Histories out of 2,880 40Atomic check: global policy · Unauthorized-effect histories: 0 Histories out of 2,880 0Atomic check: relevant policy · Unauthorized-effect histories: 0 Histories out of 2,880 0
Unauthorized effects under six executor controlsHistories out of 2,880. Histories containing at least one unauthorized effect. Each control runs on the same 2,880 constructed histories. Lower is better. Exact values follow under View data. 01,0002,000 工具允许列表Exact requestbindingValidation atpreparationCachedone-use checkAtomic check:global policyAtomic check:relevant policy 执行历史数(共 2,880 条) Tool allowlist · Unauthorized-effect histories: 2,592 Histories out of 2,880 2,592Exact request binding · Unauthorized-effect histories: 320 Histories out of 2,880 320Validation at preparation · Unauthorized-effect histories: 128 Histories out of 2,880 128Cached one-use check · Unauthorized-effect histories: 40 Histories out of 2,880 40Atomic check: global policy · Unauthorized-effect histories: 0 Histories out of 2,880 0Atomic check: relevant policy · Unauthorized-effect histories: 0 Histories out of 2,880 0
Unauthorized effects under six executor controlsHistories out of 2,880. Histories containing at least one unauthorized effect. Each control runs on the same 2,880 constructed histories. Lower is better. Exact values follow under View data. 02,000 ToolallowlistExactrequestbindingValidation atpreparationCachedone-usecheckAtomiccheck:global policyAtomiccheck:relevantpolicy Histories out of2,880 Tool allowlist · Unauthorized-effect histories: 2,592 Histories out of 2,880 2,592Exact request binding · Unauthorized-effect histories: 320 Histories out of 2,880 320Validation at preparation · Unauthorized-effect histories: 128 Histories out of 2,880 128Cached one-use check · Unauthorized-effect histories: 40 Histories out of 2,880 40Atomic check: global policy · Unauthorized-effect histories: 0 Histories out of 2,880 0Atomic check: relevant policy · Unauthorized-effect histories: 0 Histories out of 2,880 0
包含至少一次未授权执行的历史序列。各控制方案均运行相同的 2,880 条构造执行历史。数值越低越好。

Zortex Research,授权漂移研究,实验 A。

查看方法

十种请求变体 × 准备前的十二种变化 × 准备后的十二种变化 × 两种调度,为每种控制方案生成 2,880 条执行历史。独立实现的检查器将已提交的执行结果与获准的文档、接收方、上下文、策略、有效期和单次使用额度进行比较。驱动程序直接构造请求;这些计数衡量既定执行契约下的约束执行情况。

查看数据
执行历史数(共 2,880 条)
条件包含未授权执行的历史序列
工具允许列表2,592
精确请求绑定320
准备阶段校验128
缓存单次使用检查40
原子检查:全局策略0
原子检查:相关策略0

Unauthorized effects under six executor controls

包含至少一次未授权执行的历史序列。各控制方案均运行相同的 2,880 条构造执行历史。数值越低越好。

两种成功的控制方案在同一把锁内校验当前执行结果、当前授权和剩余额度,并记录额度消耗与执行结果。因此,检查的值就是提交时使用的值。

保留有效操作

拒绝所有请求的控制方案同样不会记录未授权执行。因此,我们单独统计有效执行机会:即已批准操作在提交时满足执行契约的历史序列。

全局策略版本会在任何策略变更后使批准失效,包括与操作无关的变更。它错失了 80 次有效执行机会中的 28 次。仅检查已批准操作所涉及策略的方案,错失的执行机会为 零,并同样保持零违规。该方案使用完整且显式提供的操作策略与资源依赖列表。

Safe execution must also preserve valid work

  • Completed valid work
  • Missed valid work
Safe execution must also preserve valid workValid opportunities out of 80. Each control encounters the same 80 histories in which the approved action is valid at commitment. The relevant-policy check completes all 80. Exact values follow under View data. 020406080 准备阶段校验Atomic: global policyAtomic: relevant policy Valid opportunities out of 80 Validation at preparation · Completed valid work: 72 Valid opportunities out of 80 72Validation at preparation · Missed valid work: 8 Valid opportunities out of 80 8Atomic: global policy · Completed valid work: 52 Valid opportunities out of 80 52Atomic: global policy · Missed valid work: 28 Valid opportunities out of 80 28Atomic: relevant policy · Completed valid work: 80 Valid opportunities out of 80 80Atomic: relevant policy · Missed valid work: 0 Valid opportunities out of 80 0
Safe execution must also preserve valid workValid opportunities out of 80. Each control encounters the same 80 histories in which the approved action is valid at commitment. The relevant-policy check completes all 80. Exact values follow under View data. 020406080 准备阶段校验Atomic: global policyAtomic: relevant policy Valid opportunities out of 80 Validation at preparation · Completed valid work: 72 Valid opportunities out of 80 72Validation at preparation · Missed valid work: 8 Valid opportunities out of 80 8Atomic: global policy · Completed valid work: 52 Valid opportunities out of 80 52Atomic: global policy · Missed valid work: 28 Valid opportunities out of 80 28Atomic: relevant policy · Completed valid work: 80 Valid opportunities out of 80 80Atomic: relevant policy · Missed valid work: 0 Valid opportunities out of 80 0
Safe execution must also preserve valid workValid opportunities out of 80. Each control encounters the same 80 histories in which the approved action is valid at commitment. The relevant-policy check completes all 80. Exact values follow under View data. 020406080 ValidationatpreparationAtomic:globalpolicyAtomic:relevantpolicy Valid opportunities out of 80 Validation at preparation · Completed valid work: 72 Valid opportunities out of 80 72Validation at preparation · Missed valid work: 8 Valid opportunities out of 80 8Atomic: global policy · Completed valid work: 52 Valid opportunities out of 80 52Atomic: global policy · Missed valid work: 28 Valid opportunities out of 80 28Atomic: relevant policy · Completed valid work: 80 Valid opportunities out of 80 80Atomic: relevant policy · Missed valid work: 0 Valid opportunities out of 80 0
Safe execution must also preserve valid workValid opportunities out of 80. Each control encounters the same 80 histories in which the approved action is valid at commitment. The relevant-policy check completes all 80. Exact values follow under View data. 020406080 ValidationatpreparationAtomic:globalpolicyAtomic:relevantpolicy Valid opportunities out of 80 Validation at preparation · Completed valid work: 72 Valid opportunities out of 80 72Validation at preparation · Missed valid work: 8 Valid opportunities out of 80 8Atomic: global policy · Completed valid work: 52 Valid opportunities out of 80 52Atomic: global policy · Missed valid work: 28 Valid opportunities out of 80 28Atomic: relevant policy · Completed valid work: 80 Valid opportunities out of 80 80Atomic: relevant policy · Missed valid work: 0 Valid opportunities out of 80 0
Each control encounters the same 80 histories in which the approved action is valid at commitment. The relevant-policy check completes all 80.

Zortex Research,授权漂移研究,实验 A。

查看方法

A valid opportunity is a history in which the original approval still permits the action at commitment. The independent checker identifies 80 such histories for each control. Completed and missed opportunities partition those same 80 histories. A global policy version can invalidate an approval after an unrelated change; the relevant-policy check uses the action’s declared dependencies.

查看数据
Valid opportunities out of 80
条件Completed valid workMissed valid work
准备阶段校验728
Atomic: global policy5228
Atomic: relevant policy800

Safe execution must also preserve valid work

Each control encounters the same 80 histories in which the approved action is valid at commitment. The relevant-policy check completes all 80.

反例揭示的问题

失败轨迹区分了语义变化、权限变化和重复消耗额度。每类问题都需要执行契约中的不同约束。

批准与执行发生偏离的三种方式

构造执行历史中的代表性反例。

可变状态

请求保持不变,接收方却已改变。

批准时
团队别名指向接收方 A
提交时
同一别名改为指向接收方 B
查看案例

已批准的调用指定了一个团队别名。在调用提交前,别名指向新的接收方。精确请求匹配接受了相同的文字,实际执行却将文档发送到了其他地方。

  1. 批准文档及其原始接收方。
  2. 准备未发生变化的请求。
  3. 将接收方别名改为指向其他对象。
  4. 按照新的别名值执行。

所需检查: 在提交时,将当前解析得到的接收方与获准的接收方进行比较。

撤销

较早的检查在权限失效后仍被沿用。

准备时
相关权限有效
提交时
权限已撤销,操作仍然执行
查看案例

准备阶段校验了完整操作,随后相关权限被撤销。如果提交时只检查剩余额度,仍会执行已经失去授权的操作。

  1. 在初始策略下校验操作。
  2. 撤销相关权限并更新其策略版本。
  3. 消耗尚未使用的额度,并记录执行结果。

所需检查: 在提交执行结果的同一个操作中,校验当前权限以及批准所对应的策略版本。

并发

两个调用方都看到同一项尚未使用的许可。

准备时
两个调用方都读到未使用的许可
提交时
一次已批准的效果和一次额外效果
查看案例

在任何一方消耗单次使用额度之前,两个调用方均已完成准备。如果各自沿用先前的观测,即使文档和接收方仍然正确,两者也都能执行。

  1. 第一个调用方观察到尚未使用的额度。
  2. 第二个调用方观察到同一份尚未使用的额度。
  3. 第一个调用方提交一次获准执行。
  4. 第二个调用方提交一次额外执行。

所需检查: 将当前额度检查与额度消耗原子性地结合。

跨进程边界的恢复

接收系统可能在执行器记录完成之前就已提交执行结果。如果确认消息丢失,重试必须区分“尚未发生任何操作”与“操作已经发生,但回复丢失”。单独的本地授权事务无法解决这一问题。

重试何时失效

我们使用独立的代理端和接收方进程测试了三种恢复机制,每个进程都有自己的 SQLite 数据库。五种条件覆盖正常完成及四个中断点。每种条件均进行三次使用全新数据库的重复测试,并执行一次不中断的恢复尝试。

What survives an interrupted delivery

All conditions

  • 一个批准的效果
  • 丢失的效果
  • 重复效果
What survives an interrupted deliveryShare of total (%). Five conditions, three repetitions each: 15 cases per mechanism. Segments show each outcome’s share; View data retains the case counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 12 Cases · 80.0% of 15 Cases 80%Reserve before send · Lost effect: 3 Cases · 20.0% of 15 Cases 20%Reserve before send · Duplicate effect: 0 Cases · 0.0% of 15 Cases Send before record · One approved effect: 9 Cases · 60.0% of 15 Cases 60%Send before record · Lost effect: 0 Cases · 0.0% of 15 Cases Send before record · Duplicate effect: 6 Cases · 40.0% of 15 Cases 40%Durable job + idempotent receiver · One approved effect: 15 Cases · 100.0% of 15 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 15 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 15 Cases
What survives an interrupted deliveryShare of total (%). Five conditions, three repetitions each: 15 cases per mechanism. Segments show each outcome’s share; View data retains the case counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 12 Cases · 80.0% of 15 Cases 80%Reserve before send · Lost effect: 3 Cases · 20.0% of 15 Cases 20%Reserve before send · Duplicate effect: 0 Cases · 0.0% of 15 Cases Send before record · One approved effect: 9 Cases · 60.0% of 15 Cases 60%Send before record · Lost effect: 0 Cases · 0.0% of 15 Cases Send before record · Duplicate effect: 6 Cases · 40.0% of 15 Cases 40%Durable job + idempotent receiver · One approved effect: 15 Cases · 100.0% of 15 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 15 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 15 Cases
What survives an interrupted deliveryShare of total (%). Five conditions, three repetitions each: 15 cases per mechanism. Segments show each outcome’s share; View data retains the case counts. Exact values follow under View data. 0%25%50%100% Reserve beforesendSend beforerecordDurable job +idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 12 Cases · 80.0% of 15 Cases 80%Reserve before send · Lost effect: 3 Cases · 20.0% of 15 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 15 Cases Send before record · One approved effect: 9 Cases · 60.0% of 15 Cases 60%Send before record · Lost effect: 0 Cases · 0.0% of 15 Cases Send before record · Duplicate effect: 6 Cases · 40.0% of 15 Cases 40%Durable job + idempotent receiver · One approved effect: 15 Cases · 100.0% of 15 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 15 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 15 Cases
What survives an interrupted deliveryShare of total (%). Five conditions, three repetitions each: 15 cases per mechanism. Segments show each outcome’s share; View data retains the case counts. Exact values follow under View data. 0%50%100% Reservebefore sendSend beforerecordDurable job+idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 12 Cases · 80.0% of 15 Cases 80%Reserve before send · Lost effect: 3 Cases · 20.0% of 15 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 15 Cases Send before record · One approved effect: 9 Cases · 60.0% of 15 Cases 60%Send before record · Lost effect: 0 Cases · 0.0% of 15 Cases Send before record · Duplicate effect: 6 Cases · 40.0% of 15 Cases 40%Durable job + idempotent receiver · One approved effect: 15 Cases · 100.0% of 15 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 15 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 15 Cases
Five conditions, three repetitions each: 15 cases per mechanism. Segments show each outcome’s share; View data retains the case counts.

Zortex Research,授权漂移研究,实验 B。

查看方法

Three mechanisms were tested with separate broker and receiver processes, each using its own SQLite database. Five conditions cover normal completion and four interruption points. Each condition has three fresh-database repetitions and one uninterrupted recovery attempt. All 45 planned cases completed without an unexpected failure or timeout. The selector shows either all 15 cases per mechanism or the three cases at one interruption point. The proportions describe these designed cases; they are not an estimated production failure rate. Separate SQL inspection checks the identity and number of effects in the receiver database.

查看数据
Cases原始观测值;标准化后的占比单独列出。
条件一个批准的效果One approved effect (%)丢失的效果Lost effect (%)重复效果Duplicate effect (%)
发送前预留1280.0%320.0%00.0%
记录前发送960.0%00.0%640.0%
持久化任务 + 幂等接收方15100.0%00.0%00.0%

What survives an interrupted delivery

Interruption point: All conditions

Five conditions, three repetitions each: 15 cases per mechanism. Segments show each outcome’s share; View data retains the case counts.

Normal completion

  • 一个批准的效果
  • 丢失的效果
  • 重复效果
What survives an interrupted deliveryShare of total (%). Normal completion: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). Normal completion: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). Normal completion: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%100% Reserve beforesendSend beforerecordDurable job +idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). Normal completion: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%50%100% Reservebefore sendSend beforerecordDurable job+idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
Normal completion: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

Zortex Research,授权漂移研究,实验 B。

查看方法

Three mechanisms were tested with separate broker and receiver processes, each using its own SQLite database. Five conditions cover normal completion and four interruption points. Each condition has three fresh-database repetitions and one uninterrupted recovery attempt. All 45 planned cases completed without an unexpected failure or timeout. The selector shows either all 15 cases per mechanism or the three cases at one interruption point. The proportions describe these designed cases; they are not an estimated production failure rate. Separate SQL inspection checks the identity and number of effects in the receiver database.

查看数据
Cases原始观测值;标准化后的占比单独列出。
条件一个批准的效果One approved effect (%)丢失的效果Lost effect (%)重复效果Duplicate effect (%)
发送前预留3100.0%00.0%00.0%
记录前发送3100.0%00.0%00.0%
持久化任务 + 幂等接收方3100.0%00.0%00.0%

What survives an interrupted delivery

Interruption point: Normal completion

Normal completion: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

Before delivery

  • 一个批准的效果
  • 丢失的效果
  • 重复效果
What survives an interrupted deliveryShare of total (%). Before delivery: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Lost effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). Before delivery: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Lost effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). Before delivery: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%100% Reserve beforesendSend beforerecordDurable job +idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Lost effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). Before delivery: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%50%100% Reservebefore sendSend beforerecordDurable job+idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Lost effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
Before delivery: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

Zortex Research,授权漂移研究,实验 B。

查看方法

Three mechanisms were tested with separate broker and receiver processes, each using its own SQLite database. Five conditions cover normal completion and four interruption points. Each condition has three fresh-database repetitions and one uninterrupted recovery attempt. All 45 planned cases completed without an unexpected failure or timeout. The selector shows either all 15 cases per mechanism or the three cases at one interruption point. The proportions describe these designed cases; they are not an estimated production failure rate. Separate SQL inspection checks the identity and number of effects in the receiver database.

查看数据
Cases原始观测值;标准化后的占比单独列出。
条件一个批准的效果One approved effect (%)丢失的效果Lost effect (%)重复效果Duplicate effect (%)
发送前预留00.0%3100.0%00.0%
记录前发送3100.0%00.0%00.0%
持久化任务 + 幂等接收方3100.0%00.0%00.0%

What survives an interrupted delivery

Interruption point: Before delivery

Before delivery: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

After receiver commit

  • 一个批准的效果
  • 丢失的效果
  • 重复效果
What survives an interrupted deliveryShare of total (%). After receiver commit: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After receiver commit: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After receiver commit: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%100% Reserve beforesendSend beforerecordDurable job +idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After receiver commit: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%50%100% Reservebefore sendSend beforerecordDurable job+idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
After receiver commit: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

Zortex Research,授权漂移研究,实验 B。

查看方法

Three mechanisms were tested with separate broker and receiver processes, each using its own SQLite database. Five conditions cover normal completion and four interruption points. Each condition has three fresh-database repetitions and one uninterrupted recovery attempt. All 45 planned cases completed without an unexpected failure or timeout. The selector shows either all 15 cases per mechanism or the three cases at one interruption point. The proportions describe these designed cases; they are not an estimated production failure rate. Separate SQL inspection checks the identity and number of effects in the receiver database.

查看数据
Cases原始观测值;标准化后的占比单独列出。
条件一个批准的效果One approved effect (%)丢失的效果Lost effect (%)重复效果Duplicate effect (%)
发送前预留3100.0%00.0%00.0%
记录前发送00.0%00.0%3100.0%
持久化任务 + 幂等接收方3100.0%00.0%00.0%

What survives an interrupted delivery

Interruption point: After receiver commit

After receiver commit: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

After acknowledgement

  • 一个批准的效果
  • 丢失的效果
  • 重复效果
What survives an interrupted deliveryShare of total (%). After acknowledgement: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After acknowledgement: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After acknowledgement: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%100% Reserve beforesendSend beforerecordDurable job +idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After acknowledgement: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%50%100% Reservebefore sendSend beforerecordDurable job+idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 0 Cases · 0.0% of 3 Cases Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
After acknowledgement: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

Zortex Research,授权漂移研究,实验 B。

查看方法

Three mechanisms were tested with separate broker and receiver processes, each using its own SQLite database. Five conditions cover normal completion and four interruption points. Each condition has three fresh-database repetitions and one uninterrupted recovery attempt. All 45 planned cases completed without an unexpected failure or timeout. The selector shows either all 15 cases per mechanism or the three cases at one interruption point. The proportions describe these designed cases; they are not an estimated production failure rate. Separate SQL inspection checks the identity and number of effects in the receiver database.

查看数据
Cases原始观测值;标准化后的占比单独列出。
条件一个批准的效果One approved effect (%)丢失的效果Lost effect (%)重复效果Duplicate effect (%)
发送前预留3100.0%00.0%00.0%
记录前发送00.0%00.0%3100.0%
持久化任务 + 幂等接收方3100.0%00.0%00.0%

What survives an interrupted delivery

Interruption point: After acknowledgement

After acknowledgement: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

After completion record

  • 一个批准的效果
  • 丢失的效果
  • 重复效果
What survives an interrupted deliveryShare of total (%). After completion record: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After completion record: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%75%100% 发送前预留记录前发送持久化任务 + 幂等接收方 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After completion record: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%25%50%100% Reserve beforesendSend beforerecordDurable job +idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
What survives an interrupted deliveryShare of total (%). After completion record: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts. Exact values follow under View data. 0%50%100% Reservebefore sendSend beforerecordDurable job+idempotentreceiver 总量占比(%) Reserve before send · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Reserve before send · Lost effect: 0 Cases · 0.0% of 3 Cases Reserve before send · Duplicate effect: 0 Cases · 0.0% of 3 Cases Send before record · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Send before record · Lost effect: 0 Cases · 0.0% of 3 Cases Send before record · Duplicate effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · One approved effect: 3 Cases · 100.0% of 3 Cases 100%Durable job + idempotent receiver · Lost effect: 0 Cases · 0.0% of 3 Cases Durable job + idempotent receiver · Duplicate effect: 0 Cases · 0.0% of 3 Cases
After completion record: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

Zortex Research,授权漂移研究,实验 B。

查看方法

Three mechanisms were tested with separate broker and receiver processes, each using its own SQLite database. Five conditions cover normal completion and four interruption points. Each condition has three fresh-database repetitions and one uninterrupted recovery attempt. All 45 planned cases completed without an unexpected failure or timeout. The selector shows either all 15 cases per mechanism or the three cases at one interruption point. The proportions describe these designed cases; they are not an estimated production failure rate. Separate SQL inspection checks the identity and number of effects in the receiver database.

查看数据
Cases原始观测值;标准化后的占比单独列出。
条件一个批准的效果One approved effect (%)丢失的效果Lost effect (%)重复效果Duplicate effect (%)
发送前预留3100.0%00.0%00.0%
记录前发送3100.0%00.0%00.0%
持久化任务 + 幂等接收方3100.0%00.0%00.0%

What survives an interrupted delivery

Interruption point: After completion record

After completion record: three fresh-database cases per mechanism, followed by one uninterrupted recovery attempt. Segments show outcome shares; the data table gives the original counts.

发送前预留许可在 3 个用例(共 15 个)中 出现执行遗漏,原因是执行在传递前停止。先发送、再记录完成的方案在 6 个用例(共 15 个)中 造成重复执行;这些用例中,传递已经成功,但确认或完成记录被中断。

第三种机制保存持久化待执行任务、稳定的标识符以及绑定的载荷。接收方以原子方式识别重复传递,并返回已有的回执。该契约在 全部 15 个用例中均产生了一次获准执行。接收方对标识符与载荷的持久化绑定,是被评估机制的一部分。

验证

如何核验结果

核心研究涵盖 17,280 次控制方案与执行历史的组合评估,以及 120 个强制双线程用例 和 54 个解析器用例。独立检查器直接根据执行记录重建状态变化、实际执行结果和有效执行机会。保存的重复运行结果与原结果在每条执行历史及所有线程和解析器用例上均一致。

对于恢复实验,单独的 SQL 检查核对每个接收方数据库中执行结果的标识与数量。全部 45 个设计用例 均已完成,没有发生意外失败或超时。按预期注入的进程退出仍计入观测结果。

执行契约

CaMeL 为不可信的模型计算设置了可执行的控制约束。[2] CapLease 研究持久授权与重放问题,包括接收方在防止重复外部执行中的作用。[3] Permission to Retry 将恢复过程中的当前授权、服务提供方契约和保留的观测加以区分。[4] 我们的实验通过对应的控制方案、显式状态变化以及对已提交执行结果的直接观测,分别考察这些问题。

结果指出,执行器必须连接两个边界。在本地提交时,必须在消耗额度的同时校验当前执行结果和授权。跨进程边界时,必须保留足够的状态,并获得接收方所需的保证,才能区分重复尝试与新的一次获准执行。


参考文献

  1. Derek Lilienthal and Sanghyun Hong. Mind the Gap: Time-of-Check to Time-of-Use Vulnerabilities in LLM-Enabled Agents. 2025. ↩︎

  2. Edoardo Debenedetti et al. Defeating Prompt Injections by Design. 2025. ↩︎

  3. Jinghan Xu et al. Beyond Single-Use Tokens: Durable Authorization State for Replay-Resistant LLM Agent Actions. 2026. ↩︎

  4. Deekshith Marla and Kushagra Bhatnagar. Permission to Retry: A Study of Authorization and Recovery in AI Agent Workflows. 2026. ↩︎