GitHub previews two AI scanning tools for different parts of a repository’s lifecycle

GitHub previewed AI scanning tools for different stages of the repository lifecycle.

GitHub announced two public previews on October 28 that apply AI to different security review tasks. AI Scan for repositories examines the latest commit on the default branch and reports findings in the existing code-scanning backlog. Code scanning for AI applications examines pull requests for paths where untrusted content can reach a model that has access to tools, data or business systems. A team evaluating either preview should start with the kind of code and workflow it wants to inspect. [Repository AI Scan preview](https://github.blog/changelog/2026-10-28-ai-scan-for-repositories-in-public-preview/) · [AI application scanning preview](https://github.blog/changelog/2026-10-28-code-scanning-for-ai-applications-in-public-preview/)

A repository scan starts from the default branch

After the feature is enabled, a user with write access to code scanning can start a repository scan. GitHub says an agentic scanner explores the repository within a fixed time budget and can cover languages and frameworks CodeQL does not support. Where both systems are scanning the repository, GitHub says the AI scanner skips issues CodeQL already reports. An AI Scan alert does not automatically close when a fix merges. A subsequent scan checks the new state. The preview consumes Code Security AI credits, and eligibility differs by repository ownership and visibility. [Repository AI Scan preview](https://github.blog/changelog/2026-10-28-ai-scan-for-repositories-in-public-preview/)

The AI application detector has a different starting point. It scans pull requests automatically when enabled alongside code scanning and requires GitHub Advanced Security. The product description focuses on new paths where untrusted content reaches an AI component that can take consequential actions. For example, a change could give a model a new external tool while still allowing it to read untrusted document content. A reviewer examining the resulting alert needs to understand both where the content enters and what the tool is authorized to do. [AI application scanning preview](https://github.blog/changelog/2026-10-28-code-scanning-for-ai-applications-in-public-preview/)

How to evaluate each preview

For repository scanning, start with a repository where maintainers can inspect alerts and rerun the scan after a known fix. For pull request scanning, test a representative AI application change and a case where the relevant permission or validation already prevents the unsafe action. Compare the result with source review, and record whether the alert identifies a concrete path an engineer can reproduce. Treat a clear result from one scan as evidence about the code, policy and inputs it examined.

The previews address different questions. One helps build a backlog from a repository snapshot. The other adds a review signal when a pull request changes how an AI application receives content or uses its capabilities.