安全评估 / NAVIN RESEARCH
Privacy Engineering for AI Applications
Minimize data, define retention, isolate tenants, control logs, and make deletion testable.
Minimize data, define retention, isolate tenants, control logs, and make deletion testable.
简要回答: Collect and send only data required for the task. Document where prompts, files, embeddings, caches, traces, and backups go; who can access them; how long they remain; and how deletion propagates. Provider settings do not replace application governance.
实施步骤
- Create a data-flow inventory for every model and tool call.
- Classify fields and redact before logging or evaluation storage.
- Apply tenant-aware authorization to retrieval and memory.
- Test access, export, correction, and deletion procedures.
验证清单
- Inspect logs and traces for accidental secrets.
- Verify retention jobs and backup policies.
- Review subprocessors and regional processing requirements.
常见错误
- Sending full records when a few fields suffice.
- Using production prompts as an unlabeled eval corpus.
- Claiming zero retention without checking every connected service.
实际决策
先交付能够通过上述检查的最小设计。记录测试中使用的模型、提示词、工具、数据集与运行时版本，让后续变化可以被比较而不是被猜测。依赖供应商的限制、模型名称、价格或预览功能之前，请重新检查链接文档；这些细节变化往往快于本文的工程原则。
