NewsWebsite creationAI Scan

GitHub AI Scan: check up to 20 repositories in a single run

GitHub lets you run AI Scan on 20 repositories at once. Here is how to test its usefulness, control costs and verify the alerts.

Mehdi B.

Co-founder · SEO & digital strategy

GitHub AI Scan: check up to 20 repositories in a single run
In brief

GitHub lets you run AI Scan on 20 repositories at once and groups the results with existing alerts. Test non-sensitive projects first, measure the false positives and check the budget and the conditions under which your code is processed.

On 16 November 2026, GitHub extends AI Scan to bulk launch on up to 20 repositories, the spaces that store project code. This update concerns companies that maintain several websites or applications and want to organise their security checks. We see it as a way to multiply checks, provided you measure how useful they are before rolling them out widely.

AI Scan: what changes with launching on several repositories?

Bulk launch is done from the organisation’s security alerts or when the corresponding setting is enabled. Each scan covers the default branch, the reference version of the code. The service remains in public preview, so its features may still change.

AI Scan complements CodeQL, GitHub’s code analysis tool. It looks for additional issues in the languages already analysed and also covers languages not supported by CodeQL. Its results join the same alert list, in Code Scanning.

According to GitHub, no Copilot licence is needed. Each scan does, however, consume AI credits; for an organisation's private repositories, a GitHub Code Security or Advanced Security licence is still required. Check access and the available budget before selecting your projects.

What is the benefit for a business that maintains several sites?

GitHub AI Scan: check up to 20 repositories in a single run

Take a shop whose team maintains a WordPress site, a custom PHP plugin and a Next.js application. For a first trial, select demonstration projects that are representative of these technologies. This lets you observe the results without starting with the shop's confidential code.

Grouping alerts makes them easier to assign: for each issue retained, name a person and a deadline. Also ask which sales journey is affected, for example checkout or customer account access. The severity level displayed is not enough to organise the fixes.

The analysis remains partial: the agent selects code within a limited time, and its results may be inaccurate. An empty list is therefore not a security validation. For the shop, keep a manual check on access to orders and customer data.

This approach complements the checks on the tools that host and circulate code. Our news article on Git and GitLab security covers this other side of development automation.

How do you decide whether the scanner deserves a permanent place?

Start with a small set of non-sensitive repositories. Compare the results with CodeQL, Semgrep, another code analyser, and a manual review. Keep the same version of the code for each check so that the comparison is usable.

For each alert, note whether it reveals a confirmed defect, repeats a known problem or flags a problem that does not exist. This last case is a false positive. Measure their proportion among the alerts reviewed, as well as the time spent sorting them.

Then go back to the shop: does the trial bring a useful fix on customer access or an ordering function? Compare the cost of the scan and the triage with the additional defects confirmed. Set your adoption criteria before the test, to avoid mistaking a high volume of alerts for real progress.

Which checks should you plan before the first launch?

  1. Can the code be submitted to the service? Check the conditions that apply to the processing and retention of the code and to confidentiality commitments. Have these points approved before any client project. In the demo shop, use fictitious data and remove real credentials.

  2. Which version is actually being checked? Ask the team to compare the default branch with the live version. Note the version chosen in the test report. To review the results, open “Security and quality”, then “Code scanning”, and filter on AI Scan.

  3. Who verifies the fix? After a fix is merged, rerun the scan manually: merging does not trigger a new analysis. During the preview, automatic closure depends in particular on the flagged line being modified. Have the resolution confirmed by a test of the behaviour concerned before closing the follow-up.

Sources

GitHubApplication securityArtificial intelligence
Mehdi B.
About the authorMehdi B.

Co-founder of Smart Impact. Hooked on the web from the start, he launched his first project in 2006: an online music magazine still running today. With close to 20 years of SEO experience, a Swiss federal diploma in marketing and a solid geek streak, he turns clients' ideas, sometimes hazy ones, into concrete digital projects with his team.

LinkedIn
Further reading