The problem
Prompts are unstructured, so sensitive data hides in them: a support ticket pasted into a summary, an AWS key in a stack trace, a national ID in a form. Once that reaches the provider you cannot take it back, and you often cannot even prove what left. Blanket-blocking AI is not an option, and asking every team to sanitize their own prompts does not scale.
How it works
Turn on the built-in rule packs, or write your own, and assign a policy to the traffic it should cover. A policy applies to all traffic or is scoped to a key, model, identity, tag, endpoint, tool, or agent, and each carries a mode.
- Log only records a finding and forwards the request unchanged.
- Redact replaces the value before it leaves.
- Tokenize replaces it before it leaves and restores the original in the response, so the caller sees no change and the provider never sees the value.
- Block refuses the request.
- Exempt carves out a trusted key or service from a stricter policy.
Detection runs in the gateway, on the request and on the response.
What you get
- Built-in packs for secrets and API keys, financial data, contact details, and government or national IDs, with per-country scoping.
- Checksum validation on structured data (cards, IBANs, national IDs) to keep false positives low, plus your own terms and patterns as custom packs.
- Response, tool-result, and agent-response scanning, so data cannot leak back out the other side.
- Prompt-injection and jailbreak detection on the way in.
- Findings recorded in the audit log as type and count only, never the raw value, so the evidence itself holds no sensitive data.
See governance for the full policy model.