Surfil

Block risky calls with Guard

Guard is the security stage - deliberately first in the pipeline. Every tool-call and payload is evaluated pre-flight, before anything runs: secret patterns, OWASP-scored injection, sandbox rules. Fail-closed - if policy can't be evaluated, the call doesn't run.

Prerequisite: Surfil installed and intercepting - see Install. Guard is active on the security stage from the moment interception is wired; there is no separate enable step.

When you need this

The moment you let an agent execute tool calls - shell commands, file writes, outbound requests - you have a blast-radius problem. Guard exists so a prompt-injected or simply wrong tool call gets stopped at the one interception point instead of after it ran.

What happens on each call

  1. Your agent issues a request; it crosses the single interception point.
  2. Guard evaluates it against policy pre-flight.
  3. The call is allowed, sandboxed, or blocked. If policy evaluation itself fails, the call is blocked - fail-closed, always.
  4. A Weave fact records the decision, with provenance.

What gets recorded - and what doesn't

  • Each decision becomes a signed Weave fact: what was decided, by which rule, when.
  • The blocked call's content never leaves the device - only decision metadata does. Same zero-trace rule as everything else.

How you know it worked

  • Blocked or sandboxed calls surface as decisions in your session record and as Weave facts - surfil mind stats shows the spine receiving them.
  • Allowed traffic is untouched: passthrough stays byte-exact, so normal work is indistinguishable from running without Surfil.
If a legitimate call is being blocked, that is Guard doing its job with a rule you need to adjust - fail-closed means false positives block rather than slip through. Enterprise deployments layer Policy on top of Guard to simulate a rule change against recorded traffic before it enforces anything live.
  • Weave - where Guard's decisions live.
  • Zero-trace - why blocked content never leaves the device.
  • FAQ - short answers on security posture.