GitHub Copilot security settings
GitHub Copilot Security Settings: Top 10 for Admins
GitHub Copilot comes as a personal plan (Free, Pro and Pro+), Copilot Business and Enterprise managed by your GitHub organization, and an agent on each developer's computer. These are the 10 settings that matter most, where to find each one and what good looks like.
In short: start with (1) MCP servers in Copilot policy, (2) MCP allowlist or registry-only and (3) Copilot cloud agent enablement, then work down the list below.
By Jon Rose, Founder & Managing Partner, IOmergent. Vendor facts as of October 2026. Not sure which version you have? See which AI account is which.
The top 10 GitHub Copilot security settings
Most important first. Our free AI Hardening Check walks you through these and the other 14, with direct links, inside your own AI assistant.
-
1MCP servers in Copilot policy
- Where to find it
- Organization Settings > Copilot > Policies, or enterprise AI controls > MCP: 'MCP servers in Copilot'. Editors where MCP is in preview follow 'Editor preview features'. Between orgs the least restrictive wins. Does not cover the GitHub MCP server in other hosts
- Good looks like
- Set at the enterprise and enabled only together with an allowlist (GHC-2)
- Why it matters
- Unrestricted MCP lets any server reach your code
-
2MCP allowlist or registry-only (allowlist Enterprise only)
- Where to find it
- Enterprise managed settings (copilot/managed-settings.json in .github-private, or MDM): allowedMcpServers / deniedMcpServers by serverName, serverUrl or serverCommand; GA Aug 2026. Or policy 'Restrict MCP access to registry servers' + 'MCP Registry URL' (IDEs, CLI, app; not cloud agent)
- Good looks like
- Allowlist set and registry restriction 'Allow all' so they don't conflict. Enforced by CLI, VS Code, Copilot app and JetBrains; a broken file blocks all but built-in servers
- Why it matters
- Only reviewed servers run
-
3Copilot cloud agent enablement
- Where to find it
- Organization Settings > Copilot > Cloud agent > Repository access (No repositories, All repositories, Only selected repositories). Enterprise: AI controls > Agents, 'Enabled for selected organizations'. Off by default on Business and Enterprise
- Good looks like
- Enabled for selected organizations and selected repositories only
- Why it matters
- The agent works autonomously in the repos it can reach
-
4Cloud agent firewall
- Where to find it
- Organization Settings > Copilot > Internet access: 'Enable firewall' and 'Recommended allowlist' (Enabled, Disabled, Let repositories decide; default Let repositories decide), 'Allow repository custom rules' (default Enabled), 'Organization custom allowlist'. Covers cloud agent and code review
- Good looks like
- Enable firewall and Recommended allowlist set to Enabled at the org; repository custom rules Disabled or reviewed
- Why it matters
- Limits where the agent can send data
-
5Content exclusion (Business / Enterprise)
- Where to find it
- Repository or organization settings > Copilot > Content exclusion
- Good looks like
- Secrets and sensitive paths excluded. Not honored in VS Code Edit and Agent modes, symlinks or remote filesystems, so pair it with deny rules (GHC-10)
- Why it matters
- Keeps sensitive files out of Copilot's context
-
6Partner agents (Anthropic Claude, OpenAI Codex)
- Where to find it
- Organization Settings > Copilot > Cloud agent > Partner agents toggles; enterprise in AI Controls > Agents. Separate from Copilot cloud agent: turning one off does not turn off the other
- Good looks like
- Off unless approved, and limited to the same repositories as the cloud agent
- Why it matters
- Third-party agents work in your repositories under their own vendor
-
7Bypass and permission rules in managed settings (Enterprise only)
- Where to find it
- Enterprise managed settings (copilot/managed-settings.json in .github-private, or MDM): permissions.disableBypassPermissionsMode, permissions.deny, ask, allow with Shell(), Read(), Edit(), Domain() rules. Bypass key not enforced in cloud agent
- Good looks like
- disableBypassPermissionsMode set to "disable"; deny rules for secrets files and deploy or push commands
- Why it matters
- Stops allow-all and /yolo modes in the CLI, VS Code, the Copilot app and JetBrains
-
8Copilot and agent audit log
- Where to find it
- Enterprise audit log filtered by actor:Copilot (fields actor_is_agent, agent_session_id), kept 180 days; agent sessions from the last 24 hours in AI Controls. Local prompts are not logged
- Good looks like
- Audit log streamed to your SIEM with alerts on Copilot policy changes
- Why it matters
- 180 days is the only history unless you stream it
-
9Copilot budgets that stop usage
- Where to find it
- Billing budgets for the enterprise, organization, cost center or user, with 'Stop usage when budget limit is reached'. User-level budgets always hard stop
- Good looks like
- An enterprise or org budget with stop usage on, plus a universal user-level budget
- Why it matters
- Without stop usage a budget only sends a notice and charges keep accruing
-
10Allow GitHub to use my data for AI model training
- Where to find it
- Profile picture > Copilot settings > Privacy: 'Allow GitHub to use my data for AI model training'. On by default for Free, Pro and Pro+ since April 24, 2026; earlier opt-outs kept
- Good looks like
- Disabled
- Why it matters
- Prompts, code context and file names from your sessions can train models
GitHub Copilot security settings: common questions
- What are the most important GitHub Copilot security settings?
- Start with (1) MCP servers in Copilot policy, (2) MCP allowlist or registry-only and (3) Copilot cloud agent enablement. All 10 are listed on this page, most important first, with where to find each one and what good looks like.
- How can I check all the GitHub Copilot settings quickly?
- Add the free IOmergent AI Hardening Check to your AI assistant and ask: "Run the IOmergent AI Hardening Check for our GitHub Copilot organization." It walks you through these and the other 14, with direct links, and nothing leaves your machine without your OK.
Want an expert review?
A self-check can't see which AI apps hold OAuth access to your email, files and code, or confirm the settings match what was reported. Our AI Security Assessment does both, with read-only connections and a live review.
See the AI Security Assessment