What GitHub AI Scan findings tell you about a pull request

GitHub AI Scan adds security findings to pull requests, including code that CodeQL does not cover. Those findings stay on the pull request. They do not create a full-repository scan or a security backlog, and they cannot enforce merge requirements.

GitHub's September 10 API announcement makes the scanner easier to enable across repositories. Before using those APIs, engineering leads need to understand what the resulting evidence means. An AI finding identifies a possible vulnerability in proposed code; it does not establish that the repository is secure.

The scanner examines proposed code without a build

GitHub's security detection documentation describes an engine that works directly with code in a pull request. It uses tools such as code search to gather additional repository context. A scan starts when a pull request is created and after each new commit.

AI Scan complements CodeQL, GitHub's static analysis tool. The documented coverage includes PHP, Shell/Bash, Terraform configuration written in HCL, and Dockerfiles. GitHub also names framework gaps such as JSP for Java and Blazor for C#. These are useful reasons to examine the feature when existing analysis leaves part of an application uncovered.

The detection categories describe concrete security problems. String injection includes unsafe strings used to build SQL or shell commands. Security misconfiguration includes disabled security controls and debug features. Supply chain risks include unpinned packages or images and downloads without integrity checks. These categories describe intended detection coverage, not a guarantee that the scanner finds every instance.

Although AI Scan requires CodeQL default setup, its execution does not depend on a successful CodeQL result. The AI engine needs no build and continues when CodeQL has failed or is waiting. Findings from the two engines can therefore arrive at different times. An AI result cannot substitute for the missing static analysis result.

Read the finding where GitHub displays it

AI findings appear on the Conversation and Files changed tabs, alongside CodeQL alerts. An “AI” indicator identifies their origin. Each finding contains a description of the security issue and an explanation of the risk.

Most findings include a suggested remediation, but some do not. Where a suggestion is available, Copilot Autofix provides a recommended code change. GitHub also provides thumbs-up and thumbs-down feedback for detection quality. A proposed fix still needs examination against the affected code; its presence does not demonstrate that the vulnerability has been removed.

GitHub says false positives are possible. BaristaLabs recommends that the person investigating a finding record the evidence for accepting or rejecting it on the pull request. If the issue remains open, link the follow-up work to the affected code and identify who will resolve it. This preserves the connection between the scanner's claim and the team's response.

Repository security views do not contain these findings

GitHub explicitly limits AI-powered detections to pull requests. Full-repository scans are unsupported, and the findings do not appear as backlog alerts in the repository's security view. A reviewer who checks only that view can miss an AI finding on a proposed change.

The findings are advisory: they do not block merges. They also cannot yet enforce merge requirements through rulesets, which define repository rules. Teams must retain their required security checks and decide how an unresolved AI finding affects human approval. An empty backlog or a permitted merge is not evidence that AI Scan found no vulnerability.

The APIs enable the scanner within existing prerequisites

The announcement names organization and repository endpoint paths: /orgs/{org}/code-scanning/ai-scan and /repos/{owner}/{repo}/code-scanning/ai-scan. Organization disablement takes precedence over a repository's enabled setting. Enterprise administrators must allow the feature, organizations must opt in, and repositories inherit organization enablement with an individual opt-out.

The public preview is available on GitHub.com for Advanced Security customers, without GitHub Enterprise Server support in this release. The documentation also requires a Copilot license and CodeQL default setup. Usage consumes AI credits. Consult GitHub's REST reference for request formats and authorization requirements; this article is not an API implementation guide.

AI Scan uses its own specialized prompts and ignores custom instruction files such as /.github/copilot-instructions.md and /CLAUDE.md. The separate Copilot code review default change concerns review effort, inherited effort settings, and their usage consequences. Those settings do not describe this security scanner's evidence.

For AI Scan, the useful next step is to inspect an AI-labeled finding on the pull request itself. Read the stated risk, examine any proposed fix, and record the investigation. Keep that evidence separate from CodeQL completion, repository backlog alerts, and permission to merge.

Sources

• GitHub: AI Scan for pull request APIs in public preview, September 10, 2026.
• GitHub Docs: AI-powered security detections in pull requests, captured September 11, 2026.

Product behavior is attributed to GitHub. Investigation and record-keeping recommendations are BaristaLabs guidance, not reported customer results.