Skip to main content

Safety and Data Protection

Guardrails are project policies the gateway enforces inline: on the way in (prompts) and on the way out (responses). Each policy is a collection of rules; each rule is a check type scored by a provider. About guardrails explains that model and how evaluation works; the guides that follow cover configuring guardrails and providers, redacting sensitive data, blocking prompt injection, applying stricter policy bundles to coding-assistant traffic, and the developer-side view of blocked and redacted requests.

About guardrails

The model behind the policies: rules, providers, evaluation order, and failure modes.

Configure guardrails

Create project policies and rules; validate in monitor mode before enforce.

Configure guardrail providers

Choose regexp, builtin, or tetrate for each check type.

Detect and redact sensitive data

Keep PII and credentials out of prompts and responses.

Detect and block prompt injection

Stop jailbreak patterns before they reach a model.

Protect coding-assistant traffic

Apply stricter policy bundles to agent and editor traffic without slowing normal use.

Guardrails on shared data planes

Per-project enforcement on shared capacity, gateway attachment health, and identity alerts.

Protect requests with guardrails

The developer-side view: recognising a block or a redaction and requesting coverage.