GitHub Enterprise Server 3.23 arrived as a release candidate on November 20, 2026, with several security-relevant changes. The most consequential is not one headline feature: it is the combination of local secret detection, broader malicious-package intelligence, dependency-update cooldowns and shorter-lived OAuth tokens. For enterprise platform teams, each capability changes a different part of the risk equation. None should be treated as production-ready merely because it appears in release notes.

Release candidate is not general availability

GitHub explicitly describes 3.23 as a release candidate for early testing and feedback. Its AI secret detection for air-gapped deployments and its OAuth token-expiration changes are themselves public previews. A production upgrade decision should therefore distinguish the maturity of the server release from the maturity of each included feature.

Five security tests before a rollout decision

1. Test secret detection without network access. GitHub says the new unstructured-secret detector can run entirely inside the enterprise network on the appliance CPU, without a GPU, hosted language model or outbound connection, for organizations using GitHub Secret Protection. In an isolated test environment, confirm that the detector operates with egress blocked and that logs do not expose scanned secret material. Record false positives, missed representative secrets and CPU overhead. Because this is a preview, do not assume identical behavior across future builds.

2. Treat package intelligence as a signal, not a verdict. The GitHub Advisory Database now identifies known malicious packages beyond npm across PyPI, Maven, RubyGems, NuGet, Go, Rust and Composer, using OpenSSF malicious-package advisories. Check that your dependency inventory covers each ecosystem you actually use. Verify that alerts are assigned to a remediation owner and that suppression requires a documented reason. Coverage of known packages is not a guarantee that unknown malicious packages will be found.

3. Model the three-day Dependabot delay. The release candidate makes Dependabot wait at least three days after a package version appears on its registry before opening a version-update pull request by default. This can give maintainers time to discover a compromised release, but it also introduces delay for benign updates. Build a decision rule for urgent security fixes, critical CVEs and time-sensitive compatibility changes; do not conflate a default cooldown with a complete vulnerability response policy.

4. Test OAuth refresh before shortening token lifetime. The release candidate introduces eight-hour access tokens and refresh tokens. Expiration is opt-in for existing OAuth applications and the default for newly created applications, according to GitHub. The vendor warns that clients must support requesting, storing and using refresh tokens before expiration is required. Test unattended integrations, rotated credentials, failed refreshes and revocation. A token that expires quickly can still create an outage if the refresh workflow is brittle.

5. Validate recovery, not only installation. Rehearse upgrade, rollback and backup restoration with representative repositories, authentication integrations and security workflows. Check that monitoring, audit collection and access-control behavior remain intact after a failover. Document which preview capabilities are explicitly excluded from critical production dependencies.

GHES 3.23 decision gates

Feature maturity→Security tests→Integration tests→Recovery proof

Failure at any gate means pilot only, not a production sign-off.

How to interpret the security trade-offs

These features solve different problems. Local secret detection concerns sensitive data and detection coverage. Malware advisories concern supply-chain intelligence. The Dependabot cooldown concerns timing and exposure to newly published versions. OAuth expiration concerns the duration of credential authority. Treating them as one aggregate security improvement hides operational failure modes.

For example, a dependency cooldown may reduce exposure to a newly published malicious version while increasing the time to receive a beneficial update. A short-lived token may reduce the usefulness of a stolen access token while increasing refresh-token handling complexity. The correct enterprise question is whether each control improves measured outcomes in the organization’s own environment.

Connect the upgrade to existing governance

CPX’s repository scanning release tests distinguish feature enablement from verified execution. Its secret-response controls explain why an alert is only the beginning of credential remediation. The audit evidence framework offers a complementary way to assess whether security changes leave sufficient records for independent review.

Decision record for platform leadership

Before approving the pilot, record the tested GHES build, licensing prerequisites, preview exclusions, package ecosystems, expected dependency-update latency, OAuth client compatibility and rollback owner. Use a small, representative set of repositories rather than a blanket rollout. Require the security owner and operations owner to sign the same test record so that security gains are not purchased with undocumented availability risk.

Primary evidence and limitations

The factual release details above come from GitHub’s November 20, 2026 announcement. The recommended gates are an independent operational framework. Preview behavior, supported configurations and release timing may change; validate against the current Enterprise Server release documentation before any deployment.