Privacy and data
What the API Platform sends to model providers, what Safeguard keeps about each request, and for how long.
Where your prompts go
Which provider receives a prompt is your choice, made per request: the model you name (and, with routing, the models and providers you list) decides it. The Console says so when you create a key. This is separate from Safeguard's own product features, which send no customer data to third-party AI providers.
When you call a model, the gateway sends your request to the provider that serves that model, and the provider sends back the answer. For a closed model, the provider is the company that makes it, such as OpenAI or Anthropic. For an open-weight model, it is the company that hosts it. Each model's page on safeguard.sh/models says where prompts go, under Data policy.
The gateway sends the provider your messages and the request parameters it needs. It never forwards:
- the
userfield: an end-user ID is your data, so the gateway drops it; - your API key: the gateway calls the provider with Safeguard's own credentials.
The gateway is kept apart from the rest of Safeguard. Your requests never pass through the AI provider chain behind Safeguard's own product features, and the gateway calls providers with credentials of its own. It also reads none of your workspace's security data. Automated tests enforce that boundary.
Routing
Where routing is switched on, a request can list several models and providers, or ask the gateway to sort them by price, latency or uptime. Every candidate is one you named, the same model at another host, or (with fallbacks on) a same-family fallback; provider.only and provider.ignore limit which providers can receive the prompt.
Retries and fallbacks
Off by default. If a tenant admin or super admin turns on Allow retries and fallbacks, or a request sets "provider": { "allow_fallbacks": true }, a failed request is retried once and can then go to a fallback listed in the catalog. A fallback is always a model from the same family (the same maker), or the same model at another host. So with fallbacks on, a request for an open-weight model can be served by a different host from the one you chose. It never goes to an unrelated model family.
Before a provider is used
A third-party model is served only after Safeguard has recorded three things for its provider:
- the provider's terms allow Safeguard to resell it at list price;
- the provider is on the subprocessor list on Safeguard's Trust Center;
- Safeguard's DPA, privacy notice and AI disclosure have been updated for it.
The gateway checks these on every request and refuses the model until all three are recorded. The rule covers hosts of open-weight models as well as the makers of closed ones, because both receive your prompts.
What Safeguard keeps
The request log: metadata only
Every request, from the API or the Workbench, adds one row to your workspace's request log. The row holds metadata, never the text:
- the time, the source (API or Workbench), and the API key used, with its name and masked form;
- the model, its provider and the model ID at the provider, and the model's region when the catalog sets one;
- input, output and cached tokens, and whether they are the provider's figures or an estimate;
- the cost, the latency and the time to the first token;
- the status, the HTTP status and any error code;
- whether the request streamed, was made under zero data retention, or was served by a fallback.
People in your workspace who can open the Console see the log on the Usage page, under Request log.
Prompts and outputs: only if you turn it on
Safeguard does not store the text of prompts or outputs by default. A tenant admin or super admin can turn on Store prompts and outputs in Usage › Credits and limits › Gateway settings. With it on, the full text of every prompt and every output in the workspace is stored by Safeguard.
- Stored text is not shown in the Console. Usage shows how many prompts are stored.
- Turning the setting off stops storing new requests. It does not delete text stored while it was on.
- To delete it, a tenant admin or super admin selects Delete stored prompts on Usage and confirms. Every stored prompt and output in the workspace is deleted, and this cannot be undone.
- It applies to requests made under zero data retention too. ZDR concerns the provider's copy, not Safeguard's.
API keys
Safeguard keeps a SHA-256 hash of each key, never the key itself. See Security and compliance.
No training on your data
Safeguard never uses your prompts, outputs or request logs to train models.
What a third-party provider may do with your data is set by that provider's terms. Each model's page summarizes them under Data policy › Retention. Some providers' terms allow training in some cases, for example on preview models, so read the page before you send sensitive data.
Retention
| Record | Kept by | How long |
|---|---|---|
| Request log (metadata) | Safeguard | Configurable; not yet enabled. Until Safeguard sets a period, it is kept until deleted. |
| Prompts and outputs, when Store prompts and outputs is on | Safeguard | Configurable; not yet enabled. Until Safeguard sets a period, they are kept until you delete them with Delete stored prompts. Once periods are set, a stored prompt is never kept longer than its log row. |
| Your prompts and outputs at the provider | The provider | Set by the provider's terms. See Retention and Zero data retention on the model's page. |
Once Safeguard sets the retention periods, a daily job deletes rows older than them. The credit ledger is a financial record and is not deleted. Usage shows the current retention under Retention.
To limit what providers keep, see Zero data retention. To keep no copy of the text at Safeguard, leave Store prompts and outputs off.
Legal and contacts
- Subprocessors: trust.safeguard.sh/subprocessors
- Data Processing Agreement: safeguard.sh/legal/dpa
- Privacy questions: privacy@safeguard.sh
Related
Team and access
Who can switch on the API Platform, create and manage inference keys, change gateway settings and add credit, how keys relate to organizations, and how key and billing changes are audited.
Zero data retention
How the gateway sends a request only to models whose provider offers zero data retention, and which models the catalog marks that way.