Dashboard
Guardrails
Configure enforcement, failure behaviour, blocking and masking, alerts, limits and the circuit breaker for an AI system.
Open an AI system and select Guardrails. Only admins can change these settings. Changes apply to new requests; the SDKs pick them up through the runtime configuration.
AI Firewall / Security Shield
System Purpose Context
Choose the Purpose Category (industry or domain) of the system, or Other with a custom category. It gives the detectors context about what is normal for this system.
Active Guard or Shadow Mode
| Mode | Behaviour |
|---|---|
| Active Guard | Blocking and masking are enforced before the model call. |
| Shadow Mode | Nothing is blocked or masked. Detections are logged as simulated outcomes, so you can see what Active Guard would have done. |
Use Shadow Mode to calibrate a new system on real traffic, then switch to Active Guard. In Shadow Mode the blocking and masking switches are simulated, and the system always behaves as fail-open.
Strict Security (Fail-Closed)
When on, requests are blocked if AgentID cannot complete the check (dependency or logging failure, timeout). When off, they continue (fail-open). See Failure modes.
Store Encrypted Payloads
Keeps prompts, outputs and the originals behind placeholders encrypted for forensic review. Required for the Reveal action in the activity log. See What is logged.
Core Guardrails
| Switch | Effect |
|---|---|
| Block Prompt Injection | Blocks jailbreak and prompt-override attempts. SDK / API and Hybrid systems only. |
| Block Toxicity & Profanity | Blocks hate, harassment and unsafe language. |
| Block Personal Information | Blocks requests that contain personal information before they reach the model. |
| Mask Personal Information | Replaces personal information with typed placeholders before the request leaves your side. |
| Block Secrets & API Keys | Blocks leaked passwords, bearer tokens, private keys and provider API keys. |
| Mask Secrets & API Keys | Replaces passwords, tokens and API keys with placeholders. |
| Restore masked values for the end user (Beta) | Shown when masking is on. Restores originals only on the user's device or in your application's response. |
For how blocking, masking and restoration interact, see Masking settings.
Click Save AI Firewall / Security Shield to apply.
Notification Channels
Send alerts to a Slack Incoming Webhook URL and an Alert Email. Click Save Notification Channels. For machine-readable delivery to a SIEM, use webhooks.
Service Protection
| Setting | Meaning |
|---|---|
| Max requests per minute | Per-system request rate, up to your plan maximum. |
| Daily request cap | Maximum requests per day. Leave blank to disable. The field shows when daily caps are not enforced for your system. |
| Hourly token cap / Daily token cap | Maximum tokens per hour or day. |
Requests over a limit are rejected with 429 and a reason such as rate_limited, daily_limit, hourly_token_limit or daily_token_limit. See Errors and rate limits.
Circuit Breaker
Enable Automatic Circuit Breaker stops the system automatically when errors exceed the Error Threshold (per 5 mins), and resumes it when the error rate drops. While stopped, Guard answers 503 with reason circuit_breaker.
Pause API traffic (kill switch)
The pause control in the system header stops all API traffic for the system after you confirm Pause now; resume it with the same control. A paused system rejects every request: Guard answers 403 with reason system_suspended, and the Ingest API answers 403.