Mental Model: Where Controls Live
| Layer | Examples | Applies to |
|---|---|---|
| Enterprise/Org Policies (“AI controls”) | Feature enable/disable, model access, cloud agent per-org allow | Whatever surfaces the policy covers |
managed-settings.json (file-based enterprise config) |
MCP allow/deny, marketplaces, plugins, sandbox, permissions | CLI, VS Code, Copilot app, cloud agent, JetBrains (varies by key) |
| Content controls | Content exclusion, public-code matching | Suggestion/context generation |
| Network controls | IP allow lists, subscription-based network routing, cloud agent firewall | Access to GitHub/Copilot infra, agent sandbox egress |
| Visibility/audit | Audit log, streaming, usage metrics | Governance/monitoring, not prevention |
Official documentation
Golden rule for conflicts: enterprise beats org; across multiple orgs in one enterprise, least-restrictive usually wins except for a sensitive shortlist (Metrics API, semantic indexing, public-code matching, code review without license) where most-restrictive wins; across multiple enterprises, most-restrictive wins.