加固安全护栏 · Safety Guardrail Hardener

它解决什么问题 · The problem

Agent 应用一旦接入真实业务，提示注入、越权操作和数据泄露就会变成真实风险。这个技能把「安全护栏」变成可审计、可测试、可上线的规则清单，而不是一句空泛的「请安全回答」。

Once agents touch real business systems, prompt injection, unauthorized actions, and data leaks become real risks. This skill turns safety guardrails into auditable, testable, launch-ready rules instead of a vague instruction to answer safely.

这个技能做什么 · What it does

它先盘点 Agent 的能力边界和数据权限，再针对提示注入、越权调用、敏感数据外发等风险输出分层护栏：输入过滤规则、工具白名单、输出脱敏规则，并附攻击用例和验证步骤。

It maps the agent's capabilities and data permissions first, then outputs layered guardrails for prompt injection, unauthorized tool calls, and sensitive-data exfiltration: input filters, tool allowlists, output redaction rules, plus attack cases and verification steps.

为什么比裸提示词强 · Why it beats raw prompts

• 给出可执行的拦截规则，而不是口号 / Gives enforceable blocking rules, not slogans
• 每条护栏都配攻击用例，可测试 / Every guardrail ships with attack cases, so it is testable
• 区分必须拦截、需人工确认、放行三级 / Separates must-block, human-review, and allow into three tiers
• 输出上线检查清单，防止漏配 / Outputs a go-live checklist to prevent missed configs

典型用例 · Typical use cases

• Coze / 扣子工作流上线前的安全审计 / Security audit before shipping a Coze workflow
• 客服 Agent 的敏感信息脱敏规则 / Redaction rules for customer-service agents
• 工具调用白名单与越权拦截 / Tool-call allowlists and privilege-escalation blocking
• 企业合规要求的护栏文档化 / Documenting guardrails for enterprise compliance

局限 · Limitations

输出规则、用例与清单，不修改你的线上系统；最终部署需在目标平台完成并复测。

It outputs rules, cases, and checklists but does not modify your production systems; deployment and retesting must be done on the target platform.