MERCY KEY
PREFLIGHT / MK-PF-01 / 12 CONTROLS

Free browser-only scorecard免费浏览器内评分卡

Before the agent can act, test the operating system around it.在代理可以行动前，先检验围绕它的运行体系。

Answer 12 evidence-focused questions about authority, boundaries, execution, observability, and recovery. Get a readiness signal and your first three actions.回答 12 个以证据为核心的问题，覆盖权限、边界、执行、可观测性与恢复；获得准备度信号及优先三项行动。

This is a triage tool, not a certification or safety guarantee. Score only what you can support with current evidence; treat unknown as no.这是分诊工具，不是认证或安全保证。只对有当前证据支持的项目计分；未知项按否处理。

Begin scorecard开始评分
0 of 12 answered已回答 0 / 12
01 / AuthorityDoes every agent and connected tool use a named, dedicated identity?每个代理和连接工具是否都使用明确、独立的身份？
02 / AuthorityAre permissions limited to the exact objects and actions this workflow needs?权限是否仅限于该工作流确实需要的对象与操作？
03 / AuthorityDo sending, spending, deletion, deployment, and permission changes require an appropriate approval?发送、支出、删除、部署和权限变更是否需要适当审批？Critical gate关键门槛
04 / AuthorityCan operators revoke the workflow’s credentials without relying on the agent itself?操作员能否在不依赖代理自身的情况下撤销工作流凭证？Critical gate关键门槛
05 / BoundariesIs untrusted content kept separate from operating instructions and tool authority?不可信内容是否与操作指令和工具权限保持隔离？
06 / BoundariesAre spend, volume, time, retry, and concurrency limits enforced outside the model?支出、数量、时间、重试和并发限制是否在模型之外强制执行？
07 / BoundariesAre destinations, domains, recipients, paths, and object types tightly bounded?目标、域名、收件人、路径与对象类型是否受到严格边界限制？
08 / ExecutionAre side effects idempotent and protected from ambiguous or duplicate retries?外部副作用是否具备幂等性，并能防止模糊或重复重试？
09 / EvidenceCan logs connect input, model/version, decision, approval, tool call, and downstream result?日志能否关联输入、模型/版本、决策、审批、工具调用与下游结果？
10 / EvidenceWill operators be alerted to new destinations, denied actions, repeated failures, and cost spikes?出现新目标、被拒操作、重复失败或成本激增时，操作员是否会收到告警？
11 / RecoveryDoes a tested kill switch stop active work, queues, retries, and scheduled restarts?是否有经过测试的终止开关，可停止活动任务、队列、重试和定时重启？Critical gate关键门槛
12 / RecoveryDoes recovery require a safe test, evidence review, and shadow or canary restart?恢复是否要求安全测试、证据复核以及影子或金丝雀重启？

Answer every item. “Unknown” should be scored as no.请回答全部问题；“未知”应按“否”计分。

FOUNDATION / PRIMARY GUIDANCE

A small operating check grounded in larger public frameworks.以更完整的公开框架为基础，形成精简运行检查。

NIST AI RMF OWASP Prompt Injection OWASP Excessive Agency