Data and security
Content guardrails
PII, secret, prompt-injection and safety checks you configure per identity on requests and completions.
このページはまだ翻訳されていません。英語版を表示しています。
Content guardrails inspect a chat request after authentication and before routing, and the completion before it is returned. You configure them per identity on the API keys page, Guardrails tab. They are separate from spend and rate limits, which live under Budget & limits.
Classifiers and policies
| Piece | What it is |
|---|---|
| Classifier | Where a check runs: an HTTPS endpoint you operate (JSON in, verdict out), or a keyword list |
| Policy | At most one per identity: the standard pack (seven ordered checks) or your own chain |
| Check | A stage (input or output), a capability, an action (block or redact) and a timeout |
| Protected | On a classifier timeout or failure: refuse the request, or skip the check |
A policy on the default identity governs every key created without an explicit identity. A policy can be paused without deleting it.
| At a glance | |
|---|---|
| Configured | API keys > Guardrails |
| Scope | Per identity |
| Default check timeout | 250 ms |
| Surfaces | Chat Completions, Responses, Messages |
| Content stored | None, decision counts only |
What a block looks like
An input violation never reaches a provider. An output block buffers the completion and refuses it. Redaction rewrites message or completion text; tool-call arguments are never rewritten, so a redaction on a completion with tool calls becomes a block.
| Outcome | Response |
|---|---|
| Policy violation | 400 content_filter |
| Classifier timeout or failure, protected policy | 503, retryable |
| Classifier failure, unprotected policy | Check skipped |
{
"error": {
"type": "invalid_request_error",
"code": "content_filter",
"message": "..."
}
}Coverage and cost
| Surface | Behavior under an enforcing policy |
|---|---|
| Chat Completions, Responses (HTTP and WebSocket), Messages | Inspected |
| Embeddings, images, audio, batches | Refused, because they cannot be inspected |
| Observation-only policy | Records what would happen, never refuses |
Protected checks add latency before the first byte: up to the sum of the input checks' timeouts. Prompts, completions and detector payloads are never stored; the tab shows a seven-day count of blocked, redacted and failed checks per identity.
| Your classifier endpoint | |
|---|---|
| Address | Public HTTPS only |
| Credential | Bearer token, stored encrypted, last four shown |
| Redirects, private IPs | Refused |
Not the same as content moderation
Uncensored models carry a platform moderation check that you do not configure. Guardrails are yours: you choose the checks, the classifiers and the identities they cover.
| Guardrails | Moderation | |
|---|---|---|
| Who configures | You | Experiential |
| Applies to | Your identities | Uncensored models |
| Stores content | No | No |