The problem
A tool result, a retrieved document, or a user message can carry instructions that try to override the system prompt or jailbreak the model. And a response can carry a secret or a customer record back out. Neither shows up in normal logging, and neither is something an app can catch on its own.
How it works
On the way in, the gateway scans prompts for known injection and jailbreak patterns and records or blocks them per your policy. On the way out, it scans the model response, tool results, and agent responses for secrets and PII, so the same call is covered in both directions.
What you get
- Detection of prompt-injection and jailbreak attempts, with the option to block high-severity matches.
- Output scanning on model responses, tool results, and agent responses for exfiltration.
- Findings recorded in the audit log, as type and severity, never the raw text.
- Works alongside your data-protection policies on the same call.
See governance for the policy model.