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.
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.
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
12
80.0%
3
20.0%
0
0.0%
记录前发送
9
60.0%
0
0.0%
6
40.0%
持久化任务 + 幂等接收方
15
100.0%
0
0.0%
0
0.0%
Normal completion
一个批准的效果
丢失的效果
重复效果
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
3
100.0%
0
0.0%
0
0.0%
记录前发送
3
100.0%
0
0.0%
0
0.0%
持久化任务 + 幂等接收方
3
100.0%
0
0.0%
0
0.0%
Before delivery
一个批准的效果
丢失的效果
重复效果
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
0
0.0%
3
100.0%
0
0.0%
记录前发送
3
100.0%
0
0.0%
0
0.0%
持久化任务 + 幂等接收方
3
100.0%
0
0.0%
0
0.0%
After receiver commit
一个批准的效果
丢失的效果
重复效果
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
3
100.0%
0
0.0%
0
0.0%
记录前发送
0
0.0%
0
0.0%
3
100.0%
持久化任务 + 幂等接收方
3
100.0%
0
0.0%
0
0.0%
After acknowledgement
一个批准的效果
丢失的效果
重复效果
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
3
100.0%
0
0.0%
0
0.0%
记录前发送
0
0.0%
0
0.0%
3
100.0%
持久化任务 + 幂等接收方
3
100.0%
0
0.0%
0
0.0%
After completion record
一个批准的效果
丢失的效果
重复效果
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.