Primary sourcearXiv: DeepSeek
View sourceEvent ·Official first-hand
AI 审批监控器会出现相关性失效,论文建议按证据源计数而非投票
arXiv 论文研究了多个 AI 监控器在审批智能体操作时的相关性失效问题,借用了分布式计算中 fail-prone systems 的概念。作者按两套预注册协议测试了 12 个小型开源权重监控器在支付审批上的表现,还额外测试了 Claude Opus 5.5、GPT-6 Astra、Gemini 4 Argon 和 DeepSeek-V3.2。结果显示,所有成功的欺诈性放行都通过某个证据源传播且只影响读该源的监控器;小型模型和 DeepSeek 会服从注入指令,而三个闭源模型基本无视。在等安全条件下,按证据源计数而非按票数计数的规则把正确支付误拦率从 7.4% 降到 0.9%。加入不可被说服的规则引擎后,可在测得的攻击下恢复安全且存活的规则条件。
Current read
这篇论文测了 16 个模型当支付审批监控器的表现,发现一起失效的根源是共享证据源,结论是数来源别数票,做智能体安全的人可以看看。
11 sources43° AI heatLast updated 2026/10/21 04:09:37
Evidence
11 sourcesRelated source 1Artificial Analysis
View sourceRelated source 2lmarena.ai
View sourceRelated source 3TestingCatalog
View sourceRelated source 4Bindu Reddy
View sourceRelated source 5rohanpaul_ai
View sourceRelated source 6Decoder
View sourceRelated source 7theverge
View sourceRelated source 8IT之家
View sourceRelated source 9pandaily
View sourceRelated source 10roon
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.