Platform

Platform

Guardrails

Inspect inputs, outputs, and tool calls at runtime.

Enforcement modes

Guardrails run in monitor, enforce, or strict modes. Most teams start in monitor for one sprint, then switch to enforce.

Input inspection

Detect prompt injection, jailbreak patterns, and disallowed topics before the model is called.

guardrail "input-safety" {
  applies_to = models.*
  detect [prompt_injection, off_topic("finance")]
  on_match = block
}

Output inspection

Inspect every model response. Block, redact, or escalate based on policy. Streaming responses are inspected token-buffered to catch violations before they reach the client.

Tip · Latency budget

The default output guardrail adds <15ms p95 to a streaming response.

Tool calls

Bind every tool grant to identity and purpose. The same agent calling the same tool with a different purpose is a different policy decision.

Was this page helpful?