OpenRouter 推出 Safety Retention 数据政策层级,留存仅限防滥用不用于训练
Introducing Safety Retention: A Clearer Choice for AI Privacy
- 1OpenRouter 新增 Safety Retention 数据政策层级,与 ZDR 和 Retention and Training 并列成统一框架
- 2该层级下任意请求无需被标记即可存储,仅限检测、调查和防止滥用,明确不用于模型训练
- 3防滥用系统可据此排查恶意网络活动、生化放射性核材料滥用及未经授权复制模型能力的尝试
- 4OpenRouter 与服务商可能开始按此层级留存请求,文档将更新以帮助跨服务商比较数据政策
Safety Retention 把是否留存数据与数据能否用于训练拆成两个独立问题,选不训练的层级不再等于数据不留存,企业选调用通道时两项要单独确认。
- 影响:数据敏感业务走该层级通道时,请求即便未被标记也可能被留存,只是不进训练;对客户作出不留存承诺的企业需要逐个核对所用模型对应的层级;统一三层框架降低了跨服务商比较数据政策的成本。
- 参考:可考虑先盘点现有调用各模型对应的数据政策层级,把涉及敏感数据的场景核对到 ZDR 或明确授权的层级;值得把防滥用留存写入内部数据合规清单,并与客户合同中的留存口径保持一致。
- 依据:三层定义与留存规则均出自 OpenRouter 官方公告,官方口径暂无独立验证,原文明确请求无需被标记即可存储、留存内容不用于训练。
People building with AI should be able to access advanced models and understand what happens to their data. OpenRouter’s role is to make those boundaries clear across a diverse model ecosystem.
We’re introducing Safety Retention as a new data policy tier alongside Zero Data Retention (ZDR) and Retention and Training. It allows data to be stored exclusively for abuse prevention, without being used for model training.
One consistent framework makes data policies easier to compare across models and providers. It separates the question of whether data is stored from what that data can be used for.
Three data policy tiers
Retention means keeping a copy of what you send to a model or what it sends back after processing. Our framework distinguishes three tiers:
| Tier | What it means | Is your data trained on? |
|---|---|---|
| ZDR | No routine storage after processing, as described in the ZDR policy. | No |
| Safety Retention | OpenRouter or a provider may store any request, whether flagged or not, only to detect, investigate, and prevent abuse. | No |
| Retention and Training | Data may be stored and used for training under the applicable provider policy. | May be, under the provider’s policy. |
Safety Retention limits why data can be kept. It does not require a request to be flagged before it can be stored.
Why Safety Retention exists
Safety systems can look for malicious cyber activity and misuse involving chemical, biological, radiological, or nuclear materials. They can also help detect unauthorized attempts to copy a model’s capabilities. Some threats are visible in one request; others emerge as patterns across many requests.
Retaining requests can help safety systems detect abuse before any individual request is flagged. Retained content is accessed only to detect, investigate, and prevent abuse, and is not used for model training.
OpenRouter and our providers may begin retaining requests for abuse prevention under this tier.
What this means for privacy
Storing content is a meaningful change in how data is handled. Privacy depends both on whether data is kept and on the limits placed on its use. As more capable models bring additional safeguards, those distinctions become increasingly important.
We’re updating our documentation to reflect this framework.
Safeguards and access
Broad access to advanced AI depends on safeguards providers can trust and data policies users can understand. Effective abuse prevention can help providers serve more legitimate developers and businesses, including smaller teams without direct relationships to major AI labs.
Our role is to make these differences understandable across providers, so users can compare data policies without having to learn a different privacy vocabulary each time.
Our goal is to keep advanced AI broadly accessible while making the boundaries around customer data clear.
来源:OpenRouter:Blog(Announcements) · openrouter.ai