GitHub adds a purpose-built AI model for leaked secret detection

GitHub has introduced a purpose-built AI model for leaked secret detection, expanding secret checks beyond token patterns that are easy to recognize. The company says the model reads surrounding code to identify likely credentials, including passwords that do not follow a known provider format. Existing AI-detected password alerts are being upgraded to the new model now for GitHub Secret Protection and GitHub Advanced Security customers.

The change matters because modern software teams are pushing more work through coding agents, command-line assistants and faster pull request loops. Pattern matching still catches many exposed API keys, but it struggles when a credential looks like an ordinary string, a configuration value or a test fixture. GitHub is positioning the new model as a context layer that can judge whether nearby code makes that string dangerous.

Where the model appears first

GitHub says AI-detected alert scans remain included with GHSP and GHAS at no additional charge. The company also plans to bring AI-detected alerts to GitHub Enterprise Server 3.23 in public preview, again tied to existing GHSP or GHAS coverage.

The broader rollout is more cautious. AI-detected secrets in push protection is available in private preview for eligible GitHub Team and GitHub Enterprise Cloud customers with GHSP or GHAS. When enabled, the check runs at push time so a developer can remove an unstructured credential before it lands in repository history. GitHub says an administrator must enable the feature and that organization or enterprise policies still apply.

GitHub is also adding the secret classifier to Copilot's /security-review workflow for the Copilot CLI and Copilot app. That path is planned for private preview and does not require a GHSP or GHAS license, but it will consume GitHub AI Credits once customers opt in. GitHub says the added checks are off by default and that running /security-review will not enable them on its own.

The practical tradeoff

For security teams, the useful shift is earlier detection. Catching an unstructured password before a push is cleaner than opening an incident after it appears in history, even if the repository is private. The tradeoff is operational: AI-based checks can affect billing, alert volume and developer trust if teams turn them on without clear budgets or triage rules.

CyberOGZ's read is simple: organizations should treat this as a policy feature, not only a model upgrade. Start with included alert scanning, review false positives, then pilot push protection on repositories where leaked credentials would create the most damage. The value is strongest when the model is paired with spending caps, admin controls and a workflow that tells developers exactly what to do when a suspected secret is blocked.

Sources

• github.blog

Cover photo by cottonbro studio on Pexels, used under the Pexels License.