Policy-Based Controls for Agents
Not every agent action is dangerous. Reading a README file is fine. Reading your SSH private key is not. The challenge is drawing the line — and doing it consistently.
Policies over heuristics
Runtime Guard uses policy-based controls rather than pure heuristics. This means you define rules:
- Allow: Agent can perform this action without restriction
- Block: Action is prevented and logged
- Approve: Action is paused until you explicitly approve it
Example policies
| Rule | Tool | Target | Action |
|---|---|---|---|
| Block SSH key access | filesystem.read | ~/.ssh/* | Block |
| Approve outbound to unknown domains | network.request | !allowlist | Approve |
| Allow project file reads | filesystem.read | ./project/* | Allow |
| Block startup folder writes | filesystem.write | ~/.config/autostart/* | Block |
Why policies work
- Predictable: You know exactly what will be blocked before it happens
- Auditable: Every policy decision is logged with context
- Customizable: Different workflows need different rules
What's coming
Policy templates (Starter, Strict, Crypto, Dev) are planned for v0.3. These give you a starting point you can customize for your specific workflow.
See policies in action: run a demo scan with Strict or Balanced mode.
Try Runtime Guard
See runtime security in action or request early access.