本文结构In This Post
一、为什么归因是刚需I. Why Attribution Is Mandatory
在传统系统中，日志往往只记录 API 调用结果。AI 场景下还需要记录“提示词上下文”和“触发主体链路”，否则很难判断行为意图与责任归属。Traditional logs often capture API outcomes only. In AI workloads, you must also capture prompt context and principal chain to determine intent and accountability.
二、归因数据模型II. Attribution Data Model
建议至少记录以下字段：最终触发用户、代执行 Agent、会话 ID、策略决策 ID、调用目标、结果状态、时间戳。必要时可附带 Prompt 哈希而非明文，平衡隐私与可审计性。At minimum, record triggering user, acting agent, session ID, policy decision ID, target system, outcome, and timestamp. For privacy, store prompt hash instead of full plaintext where appropriate.
三、审计链设计原则III. Audit Chain Design Principles
不可篡改Tamper-Resistant
关键审计记录应具备签名或哈希链机制，防止事后修改。Critical records should be protected with signatures or hash chaining to prevent post-event tampering.
可关联Correlatable
一次 AI 请求通常跨多个系统，日志必须可关联到同一 trace ID。AI requests span multiple systems; logs must correlate through a shared trace ID.
可导出Exportable
合规审计常要求外部证据，日志导出格式需稳定且可验证。Compliance audits often require external evidence, so export format must remain stable and verifiable.
四、事件响应场景IV. Incident Response Use Cases
一旦出现异常调用，归因链可以快速回答：是哪个用户授权、哪个 Agent 执行、在哪个策略下放行、影响了哪些资源。响应速度和结论可信度都会显著提升。During incidents, attribution chains quickly reveal who authorized, which agent acted, under which policy decision, and which resources were affected.