Arcveil 用 Jev 做结构化判定防提示注入

Arcveil @Arcveil_AI原推发布于 2026.09.20

在 X 上打开原推

Prompt injection works because the model can write the next instruction. A structured evaluation model cannot. Jev's only output is typed a probability, a label, a point on a scale. Text that tries to hijack an agent can move the number. It cannot become a command. That is why the judgement clauses in an Arcveil mandate go to Jev and not to a chat model. Built and tested

这条是引用着另一条写的：「x402 客户端用 Jev 做工具选择打分」