Jev 夹在工具调用与真实执行之间做风险闸门

Eli @ellishlomo原推发布于 2026.09.24

在 X 上打开原推

Jev can sit between a proposed tool call and real execution. • Split real intent from untrusted web junk first • The agent only proposes. Never runs • Sanitize and verify authorization and validate every argument • Jev scores risk and uncertainty. Not chat • Deterministic policy: allow, block, or send to review • Scoped tools run last, least privilege The result is minimal hallucinations.