Skip to content

Risk Scope ​

risk_scope 描述候选动作触达哪类风险边界,让 policy、UI、runtime、evidence 使用同一套风险语言。

Scope types ​

tool_call、model_call、artifact_export、source_access、credential_use、filesystem、network、remote_agent、data_retention、redaction、human_action、memory_write、task_delegation、ui_surface、cost_or_quota。

Risk levels ​

low 通常可自动允许并审计;medium 应可见且受约束;high 通常需要审批、沙箱或窄 grant;critical 需要明确权威和强 evidence;unknown 应先阻断。

Draft standard for portable agent policy decisions, approvals, permissions, risk, and audit traces.