ZeroHuman
◈

Policies

policy.yaml

Define enforceable rules: permissions, approvals, network policy, data sensitivity.

Policies are the governance layer. They map to runtime guardrails, tool approvals, and network restrictions. Policy enforces "most restrictive wins" — lower scopes cannot broaden permissions. Policies define data sensitivity classifications, approval gates, allow/deny lists for tool operations, network egress controls, and observability settings.

Schema

policy.yamlyaml
# policy.yaml — API Security Policy

name: api-security
version: "2.0"
scope: organization
severity: critical

rules:
  - name: no-hardcoded-secrets
    description: "Never hardcode API keys, passwords, or tokens"
    enforcement: block
    check: "No string literals matching secret patterns"

  - name: input-validation
    description: "All API inputs must be validated"
    enforcement: warn
    check: "Every endpoint has input validation middleware"

  - name: rate-limiting
    description: "All public endpoints must have rate limiting"
    enforcement: block
    check: "Rate limiter middleware present on all routes"

  - name: sql-injection
    description: "Use parameterized queries only"
    enforcement: block
    check: "No raw SQL string concatenation"

  - name: auth-required
    description: "All non-public endpoints require authentication"
    enforcement: block
    check: "Auth middleware on all protected routes"

approval_gates:
  - name: security-review
    required_for:
      - "Changes to authentication logic"
      - "New API endpoints"
      - "Database schema changes"
    approvers:
      - security-team
      - tech-lead

Usage

# Policies apply at the organization level
# They override all other protocol objects

# A policy can reference specific tools it governs:
applies_to:
  tools:
    - send-email
    - execute-query
    - deploy-production

Best Practices

  • ✓Scope policies clearly — organization, team, or project level
  • ✓Use "block" enforcement for critical security rules
  • ✓Use "warn" enforcement for best practices that allow exceptions
  • ✓Define approval gates for high-risk operations
  • ✓Version policies and maintain a changelog
  • ✓Keep rules testable — vague rules cannot be enforced

Anti-Patterns

  • ✗Policies so strict that agents cannot complete any work
  • ✗Unversioned policies that change without notice
  • ✗Rules without clear enforcement levels
  • ✗Missing approval gates for destructive operations