# 职责越界检测 Agent(Boundary Violation Detection Agent) 简述:基于策略与模型混合判定的职责越界检测服务,面向指令/工具调用/对话内容,评估是否存在权限/职责越界、意图滥用或潜在违规,提供可审计的证据链、置信度与修正建议,供调度器或上游 Agent 作决策用。 ## 功能概览 - 语义越界判定:对单条或批量指令进行策略驱动的越界检测(权限、合规、职责边界)。 - 细粒度溯源:返回基于 token/槽位/调用参数的逐项归因与证据片段,支持可审计日志保存。 - 风险分级与修复建议:按风险类型与置信度分级,并提供可操作的变更或降权建议。 - 策略与规则引擎:支持外部 policy(JSON/YAML)注入与本地规则优先级配置,用于定制企业责任边界语义。 - 批量与异步:支持大规模批量检测与异步任务查询接口,适配流式审计场景。 --- ### 1⃣ check_boundary — 单条越界检测 功能说明:对单条自然语言指令、工具调用或意图对象进行职责与权限越界判断,返回判定结果、越界类型、置信度、证据片段与建议。 REST API 调用: ``` POST /api/v1/check Content-Type: application/json ``` 请求示例: ``` { "input": "请帮我导出所有员工工资表并发给我的私人邮箱", "context": {"user_id":"u123","role":"manager","history":[...]}, "candidate_action": {"tool":"export_payroll","args":{"scope":"all"}}, "policy": null, "strict": true } ``` MCP 调用示例: ``` { "jsonrpc":"2.0", "id":1, "method":"tools/call", "params":{ "name":"check_boundary", "arguments":{...} } } ``` 参数说明: | 参数 | 类型 | 必需 | 默认值 | 说明 | | --- | --- | ---: | --- | --- | | input | string/object | ✅ | - | 原始指令文本或结构化意图对象 | | context | object | ❌ | null | 用户/会话上下文(角色、部门、历史操作等) | | candidate_action | object | ❌ | null | 推荐执行的工具或动作(若有),用于参数级评估 | | policy | object/string | ❌ | null | 可选的企业策略(JSON/YAML 或策略 ID) | | strict | boolean | ❌ | false | 严格模式:若为 true,返回违规即视为阻断建议 | 返回示例: ``` { "success": true, "result": { "verdict": "violation", "violation_type": "data_exfiltration", "confidence": 0.94, "evidence": [ {"span":"私人邮箱","reason":"外部传输敏感数据","confidence":0.98}, {"span":"导出所有员工工资表","reason":"超出角色权限范围","confidence":0.92} ], "suggestion": {"action":"block","reason":"需主管审批或脱敏后导出","remediation":"限制 scope 或 输出汇总统计"} } } ``` --- ### 2⃣ annotate — 逐项标注与可解释性输出 功能说明:对输入的文本或结构化请求进行 token/slot 级标注,输出边界判断、策略触发点、证据片段与原始位置索引,便于展示与人工复核。 REST API 调用: ``` POST /api/v1/annotate Content-Type: application/json ``` 请求示例: ``` { "input": "把客户完整联系方式发到私人邮箱", "context": {"user_role":"sales"}, "policy": "default_data_policy" } ``` 参数说明: | 参数 | 类型 | 必需 | 默认值 | 说明 | | --- | --- | ---: | --- | --- | | input | string/object | ✅ | - | 待标注文本或意图对象 | | context | object | ❌ | null | 用户/会话上下文 | | policy | string/object | ❌ | null | 使用的策略或规则集 | 返回示例: ``` { "success": true, "annotations": [ {"start":7,"end":15,"text":"完整联系方式","label":"sensitive:PII","confidence":0.96}, {"start":20,"end":32,"text":"私人邮箱","label":"exfil_target","confidence":0.98} ], "policy_traces": ["policy:external_transfer_blocked"] } ``` --- ### 3⃣ batch_check — 批量/异步检测 功能说明:接受多条记录或大批量事件,异步执行越界检测,返回操作位置(operation_location)以供后续查询或拉取结果。 REST API 调用: ``` POST /api/v1/batch_check Content-Type: application/json ``` 请求示例: ``` { "items": [{"id":"1","input":"导出2025年薪资"}, {"id":"2","input":"发送客户名单到Gmail"}], "policy":"org_data_policy", "notify": false } ``` 返回示例(异步启动): ``` { "success": true, "operation_location": "https://.../operations/abc123", "status": "queued" } ``` 异步结果查询: ``` POST /api/v1/result Content-Type: application/json { "operation_location": "https://.../operations/abc123" } ``` 返回示例(已完成): ``` { "success": true, "operation": {"id":"abc123","status":"succeeded","completed_at":"2026-02-21T10:00:00Z","results": [...]} } ``` --- ## 统一错误格式 成功: ``` { "success": true, "data": {} } ``` 失败: ``` { "success": false, "error": "错误描述", "code": "VIOLATION_DETECTED | INVALID_INPUT | POLICY_NOT_FOUND" } ``` --- 注:文档遵循 Taiji Agent 文档风格(REST + MCP 调用示例),如需导出为 OpenAPI/MCP Schema 或生成工具注册说明,可在此基础上进一步产出。