用 Jev 给 Agent 加一层防诈骗判断

timour kosters @timourxyz原推发布于 2026.09.18

在 X 上打开原推

Playing with Jev by @typesafeai, and it's pretty cool to think about the use cases. One idea: It may allow us to trust our agents to be much more agentic online. For every decision my agent makes that could be sensitive, Jev has to compute the probability that it's getting scammed or prompt-injected, and stop it from doing anything crazy. Kind of like the Sentinel feature from @Muse, but you can edit it and use it in many contexts.