A secure server is illustrated with glowing network, access-control, monitoring, and data-storage interfaces.
GitHub Enterprise Server 3.23 has reached release candidate status. It adds on-appliance AI secret detection, wider malicious-package advisories, shorter-lived OAuth tokens and a new way to isolate Redis in a cluster. It is a test build, not something to put on your production appliance.

A secure server is illustrated with glowing network, access-control, monitoring, and data-storage interfaces. Release status and the rules for RC builds​

GitHub's release notes label the build 3.23.0-rc.1, dated November 19, 2026. The release table has no general-availability date yet. That means 3.23 is still a candidate, even though the GitHub changelog opens by saying it is "now available."

GitHub's own warnings in the release notes are blunt:

  • RC builds are intended solely for use in a test environment. Don't install one in production.
  • Don't upgrade to an RC from a supported earlier version.
  • An instance running an RC can't be upgraded to the GA release, and it can't take a hotpatch. The usual advice is to destroy the test environment when you're done.
  • The RC doesn't include every security fix that will ship in GA. That includes some fixes already released in patches for earlier versions.

So treat 3.23 as a staging experiment. Don't treat it as a patch candidate.

What's new in GHES 3.23​

Projects filters with OR, AND and parentheses​

Project views can now combine filters with logical OR, explicit AND and parentheses for grouping. GitHub's example is (status:Done OR status:Todo) is:issue, which returns issues in either status. The announcement doesn't give further syntax limits, so test your more elaborate queries in staging.

AI secret detection that works air-gapped (public preview)​

This feature is for enterprises that use GitHub Secret Protection. It finds unstructured secrets, such as passwords and other tokens that don't follow a fixed pattern, without leaving your network. GitHub says a self-contained, fine-tuned model runs on the appliance CPU. It needs no GPU, no hosted LLM and no outbound connection. The release notes file it under GitHub Advanced Security, and it is in public preview and may change.

For regulated or disconnected environments, this addresses a real gap. Cloud-only AI scanning has been unavailable to them. The trade-off is that a CPU-hosted model shares the appliance's resources. GitHub hasn't published capacity figures in the material I reviewed, so watch performance during your trial.

Malicious-package advisories beyond npm​

The GitHub Advisory Database now identifies known malicious packages in PyPI, Maven, RubyGems, NuGet, Go, Rust and Composer, in addition to npm. It ingests malware advisories automatically from OpenSSF's malicious-packages repository.

This is coverage of known malicious packages. It isn't a guarantee that every bad package will be caught. NuGet and Maven matter here for Windows and .NET shops, since many enterprise build pipelines depend on them.

Dependabot's three-day cooldown​

By default, Dependabot now waits until a package version has been on its registry for at least three days before opening a version update pull request. The idea is to give the community time to spot and report compromised or buggy releases before anyone merges them. The announcement describes version updates specifically. It doesn't say the cooldown applies to security updates.

If your team expects same-day dependency PRs, this will change your workflow. Most teams will probably accept the trade for the extra safety margin.

Expiring OAuth tokens (public preview)​

OAuth app access tokens can now expire after eight hours, with refresh tokens used to get new ones. Expiration is opt-in for existing apps and the default for newly created apps. The feature is in public preview and may change.

  • To test or roll out gradually, request expiring tokens at authorization with scope=offline_access.
  • To enforce expiration, require it in the app registration.
  • Before you require it, update every client to request, store and use refresh tokens. Clients that can't handle them will have their sign-in break.

Inventory your internal OAuth apps and integrations first. A script or desktop tool that stores a single token forever is exactly the kind of client that will fail.

Moving the Redis primary to a dedicated cluster node​

In a clustered deployment, administrators can move the Redis primary off a node that also runs services such as MySQL. Redis then uses the dedicated node's resources and data volume. GitHub notes that this doesn't attach a new disk to an existing node and doesn't use an external Redis service. The documentation says GitHub determines clustering eligibility, and your license must enable it.

GitHub's procedure, as I read it in the documentation, has these elements:

  1. Prerequisites. Schedule a maintenance window and take a verified, recent backup. Every node, including the new one, must run the exact same GHES release. The new node needs a unique hostname, enough memory for the Redis dataset and a separate data volume with free space. You must also keep at least one healthy Redis replica besides the primary.
  2. Add the node as a replica first. Edit /data/user/common/cluster.conf on one retained cluster node, with no MySQL or other service roles assigned. Then use ghe-cluster-config-init and ghe-cluster-config-apply. Applying the configuration restarts Redis on all Redis nodes, including the current primary.
  3. Confirm health. ghe-cluster-status-redis -v should report ok and replication is in sync for the new node, and ok for the current primary. An unhealthy node is a stop condition. Before promotion, also check replication status and offsets rather than trusting the general "in sync" label.
  4. Promote in the maintenance window. Enable maintenance mode and change only redis-master in the [cluster] section. Validate, then apply.
  5. Verify afterward. Check Redis routing and replication, and confirm the MySQL primary didn't move. Disable maintenance mode, then check user access and that background jobs resume.

Two warnings stand out. First, promoting an out-of-date or empty replica can delete Redis data, including background-job queue state. Second, GitHub says the procedure doesn't guarantee zero data loss. Writes made during the primary change can be lost even if the replica was in sync. GitHub provides a way back, because the original node stays on as a replica.

Other notable items in the 3.23.0-rc.1 notes​

  • You can optionally configure a dedicated MySQL disk during a new installation. Clusters can also configure a separate dedicated data disk for MySQL.
  • Copilot CLI support for GHES in disconnected or air-gapped environments is generally available. An administrator configures a model provider once, and users sign in with their GHES credentials.
  • The bundled CodeQL CLI is 2.27.0. GitHub lists support for Go 1.27, Kotlin 2.4.10, Swift 6.3.3 and native Linux ARM64.
  • Secret scanning gains new detections and push-protection coverage, and code scanning gains a "Mitigated" dismissal reason.
  • The GHES 3.22 release notes say the Node Eligibility Service for cluster topologies will be removed in GitHub Enterprise Server 3.23. They add that Collectd metrics will be retired starting in version 3.23, with no Collectd stack in 3.23. If you rely on Collectd dashboards or alerts, plan a replacement before upgrading.

Planning notes for administrators​

  • GHES 3.22 became generally available on September 8, 2026. GitHub supports at least the four most recent feature releases, so there's no need to rush from 3.22.
  • GitHub lists CodeQL CLI 2.27.0 as recommended for 3.23. The minimum Actions runner version is 2.337.0. Most instances update runners automatically. If you use ephemeral self-hosted runners with automatic updates disabled, upgrade the runners before upgrading GHES.
  • In a staging environment, test three things first: the OAuth refresh-token behavior in your own apps, the AI secret detection load on the appliance CPU, and any dependency on Collectd.

The bottom line​

The 3.23 candidate is strongest on security. Air-gapped AI secret detection, broader malware advisories and the Dependabot cooldown all aim at supply-chain risk, and expiring OAuth tokens shrink the damage a stolen token can do. Two of the headline items, AI secret detection and OAuth expiration, are still public preview. The Redis move is useful for busy clusters but carries a real data-loss caveat. Try it all in a throwaway staging instance, send feedback to GitHub Support, and wait for GA before touching production.

 

References

  1. GitHub Enterprise Server 3.23 release candidate GitHub Changelog 2026-11-20T20:41:26+00:00
  2. (product-watch) Enterprise Server 3.23.0-rc.1 · Issue #291 · shariqh/github-enterprise-settings-configurator · GitHub github.com
  3. About upgrades to new releases - GitHub Enterprise Server 3.23 Docs docs.github.com