AI Gateway evaluates policies in three sections, in order: Access Control, then DLP, then AI Security Guardrails. Within each section, policies are evaluated top-down, and the first match wins. If no policy matches anywhere in the sequence, AI Gateway currently defaults to allowing the traffic.
AI Gateway policies support the following actions:
| Action | What it does | Stops evaluation of remaining policies? |
|---|---|---|
| Allow | Permits the traffic. | Yes |
| Monitor | Logs the activity for visibility and passes the request on to the next policy. | No |
| Block | Terminates the connection and drops the traffic. | Yes |
| Replace | Terminates the connection and responds with the replaced message. | Yes |
Use the Allow action when trusted, well-understood traffic should proceed without additional policy checks. When traffic matches an Allow policy, AI Gateway permits the request, stops evaluating remaining policies in the Access Control → DLP → AI Security Guardrails sequence for that event, and records the match in event logs for visibility. Prompts, responses, and tool or MCP events are evaluated independently, so an Allow match applies only to the specific event that matched; other events in the same session continue to be evaluated against your policies.
Things to know
- Traffic permitted by this action is still recorded and remains visible in AI Gateway event logs for auditing.
- Because Allow is terminating, verify policy order after adding an Allow rule to avoid unintentionally bypassing required checks.