Policy registry

HOL Guard policy recipes

Twenty reviewed starting points for common agent-security decisions. Each recipe uses a narrow matcher, includes safe test cases and limitations, and opens the existing Policy Studio for review before rollout.

A recipe is not an automatically enforced rule. Applying a reviewed policy through Guard Cloud uses the existing Guard policy-bundle delivery path, including RSA-PSS signing when policy-bundle signing is configured.

GPR-001 · critical

review

Review secret-file reads

Require review before supported agents read common local secret files.

Matcher: path = .env

GPR-002 · high

review

Review untrusted package installs

Require review before supported package-install actions use an untrusted package.

Matcher: package = untrusted-package

GPR-003 · critical

block

Block known-malicious package installs

Block supported install actions for packages already classified as malicious.

Matcher: package = malicious-package

GPR-004 · high

review

Review remote install scripts

Require review before a supported shell action invokes a remote installer pattern.

Matcher: command = curl

GPR-005 · high

review

Review new MCP servers

Require review when a supported harness attempts to use an MCP server that has not been reviewed.

Matcher: mcp = unreviewed-mcp-server

GPR-006 · critical

review

Review destructive shell commands

Require review before supported shell execution reaches a destructive command family.

Matcher: command = rm

GPR-007 · low

allow

Allow a reviewed domain

Allow network actions to a domain that your team has explicitly reviewed and approved.

Matcher: domain = trusted.example.com

GPR-008 · medium

allow

Allow a reviewed skill

Allow a specific skill identifier after its source and behavior have been reviewed.

Matcher: tool = trusted-skill

GPR-009 · high

review

Review high-risk tool actions

Route a named high-risk tool action to human review before execution.

Matcher: tool = high-risk-action

GPR-010 · high

block

Block a dangerous tool

Block a named tool after a security review determines that it should not execute.

Matcher: tool = dangerous-tool

GPR-011 · critical

review

Review SSH key reads

Require review before supported file-read surfaces access a common SSH private-key path.

Matcher: path = .ssh/id_rsa

GPR-012 · critical

review

Review cloud credential reads

Require review before supported file-read surfaces access a common cloud credential path.

Matcher: path = .aws/credentials

GPR-013 · high

review

Review package-manager configuration access

Require review when supported file-read surfaces access package-manager configuration that can redirect installs.

Matcher: path = .npmrc

GPR-014 · critical

review

Review Git credential access

Require review before supported file-read surfaces access a Git credential store.

Matcher: path = .git-credentials

GPR-015 · high

block

Block production package installs

Provide a conservative starting point for blocking a reviewed production package-install target.

Matcher: package = production-package

GPR-016 · high

review

Review remote MCP connections

Require review before using a named remote MCP endpoint or server identity.

Matcher: mcp = remote-mcp-server

GPR-017 · high

review

Review browser automation tools

Require review before a named browser automation tool performs a supported action.

Matcher: tool = browser-automation

GPR-018 · critical

block

Block a reviewed destructive command

Block a specific destructive command family after confirming it is unsafe for the target workspace.

Matcher: command = mkfs

GPR-019 · high

review

Review plugin execution

Require review before a named unreviewed plugin or skill-like tool executes.

Matcher: tool = unreviewed-plugin

GPR-020 · low

allow

Allow a reviewed workflow

Allow a named routine workflow only after its actions have been reviewed and bounded.

Matcher: tool = trusted-workflow