Prompt Security
What Privian guarantees about prompt handling inside the gateway.
Guarantees
- Masked egress. No prompt leaves the gateway toward a provider before the masking pipeline runs.
- No raw-prompt logging. Raw prompts exist only in request memory and never appear in logs, telemetry, response metadata, or the observability outbox.
- Sanitized observability. All event envelopes pass through a sanitizer that strips strings matching email/phone shapes and clips long strings before any sink receives them.
- Bounded fallback. The optional LLM fallback runs at most once per request with a strict timeout, untrusted-data sentinels in the system prompt, and a defensive JSON parser. It can only supplement deterministic entities, never overwrite them.
- Pre-provider validation. Auth, schema, model allowlist, quota, and rate limit checks all fail before any provider call.
What it is not
- Not a prompt-injection firewall for your application's own LLM agent.
- Not a jailbreak detector.
- Not a content-policy enforcer.