GitHub’s November 16, 2026 enterprise customization controls introduce a useful distinction: deciding where a coding assistant may obtain executable extensions is not the same as deciding which extensions are safe. Enterprises that collapse those questions into a single “plugin-only” policy can create a misleading sense of protection.
What the new control actually does
GitHub says the strictPluginOnlyCustomization managed setting can block standalone user- and project-sourced customizations while permitting eligible plugins and managed settings. Administrators can apply it independently to five categories: skills, custom agents, hooks, MCP servers, and executable Copilot extensions. GitHub explicitly cautions that plugin-only does not mean enterprise-managed plugins only and does not itself enforce a marketplace allowlist. Its separate strictKnownMarketplaces control governs eligible marketplace sources. These capabilities are described for GitHub Enterprise Cloud members whose Copilot Business or Enterprise licenses were issued by the enterprise or its organizations.
The three independent decisions
For a governed deployment, administrators need three different answers. Source policy: must a customization arrive through a plugin or managed setting? Distribution policy: which marketplaces and packages may supply it? Execution policy: what credentials, repositories, networks, and data may it touch once installed? Passing the first two does not establish the third. An approved package can still be overprivileged or updated after its initial review.
Plugin or managed setting2. Supplier approved
Marketplace and package3. Runtime bounded
Identity, egress, audit
Editorial control model; not a diagram supplied by GitHub.
Six tests before broad rollout
1. Standalone bypass: attempt a user-local MCP server and a project-local executable extension in a test organization. Confirm the selected categories are refused while intentionally unrestricted categories still function. Record the managed policy revision.
2. Marketplace provenance: test a plugin from an unapproved marketplace. Confirm that marketplace restrictions, not merely plugin-only restrictions, deny installation. Verify the exact package version and publisher identity rather than treating a familiar display name as evidence.
3. Managed-setting precedence: test conflicting user, repository, and enterprise configuration. Capture the effective setting and confirm the enterprise requirement wins in each supported client context. Do not assume policy coverage extends to unlicensed or unmanaged environments.
4. Credential scope: install an approved MCP tool with a deliberately restricted identity. Try accessing an unauthorized repository and a restricted secret. A package’s approval must not grant additional rights to the caller.
5. Update control: simulate a new plugin version that adds a network destination or tool capability. Require review of changed permissions, an owner, and rollback evidence. The important unit of approval is a specific release and behavior, not a perpetual package name.
6. Incident revocation: disable a compromised customization and verify its tool calls stop, associated tokens are invalidated where necessary, and relevant audit records remain accessible. Time the interval between administrative revocation and actual execution denial.
Make enforcement observable
Keep a small inventory joining customization name, version, origin, marketplace, responsible owner, allowed tools, credentials, and last review. Track unauthorized-install attempts separately from unauthorized-tool-call attempts. The first measures distribution enforcement; the second measures runtime boundaries. A low installation-violation rate is not evidence of low privilege.
These tests complement CPX’s analysis of OAuth delegation boundaries for agent tools, code-change governance in Slack and Teams, and authorization tests for agent-triggered workflows. The broader AI Risk Assessment research hub provides the appropriate governance context.
Procurement and security implications
Request from the platform owner an enforceable inventory of allowed sources, approved package versions, documented policy exceptions, and test results for both package installation and tool execution. Avoid presenting a configuration flag as proof of enterprise-wide isolation. An organization with multiple development environments also needs to identify which clients and licenses are in scope and which remain outside the policy.
The practical decision is not whether plugins are safer than standalone files in the abstract. It is whether the enterprise can prove provenance, constrain authority, observe changes, and revoke execution. Each of those claims requires different evidence.
Primary source
GitHub Changelog, November 16, 2026: Control where GitHub Copilot customizations come from. Product-specific claims above are attributed to GitHub; the release tests and governance model are independent editorial analysis.
Leave A Comment