v1.7.20

CyberStrikeAI 现已接入 TypeSafe Jev 作为人机协同审批引擎：Agent 每次非白名单工具调用，会把工具名、参数和策略一次性交给 Jev 做结构化放通/拦截，不再依赖聊天模型输出 JSON。Jev 并行判断破坏性风险与授权渗透意图，本地规则合成 approve/reject；默认放行授权渗透，只拦截打挂业务、毁数据、改账号权限或改线上配置的操作。自定义审批策略会作为组织规则叠加，OpenAI 协议小模型仍可选用。

---

CyberStrikeAI now supports TypeSafe Jev as a HITL audit backend. Each non-allowlisted tool call is sent to Jev in one System One request, which returns structured pass/block scores instead of chat JSON. Built-in rules approve authorized pentest activity and reject live-system destruction, irreversible data loss, account/ACL changes, and config tampering. Custom org policy is evaluated as an overlay; the OpenAI-compatible reviewer remains available.