What changed in security overview
According to GitHub's changelog, AI Scan produces additional security findings with an AI-based engine that runs on pull requests and complements CodeQL. The new reporting tracks that activity at organization and enterprise level. GitHub says the metrics show:
- how many AI Scan findings were generated and resolved
- the remediation rate
- which repositories produced findings
- trends over the selected reporting period
- how often developers applied autofix suggestions
There is also a navigation change. The old "CodeQL pull requests" page is now called "Code scanning on pull requests". It has two tabs, CodeQL and AI Scan. The metrics that used to sit on the CodeQL pull-requests page now live under the CodeQL tab, so existing bookmarks and habits may need a small adjustment. The AI Scan tab shows total findings and splits them into findings that include an autofix suggestion and findings that don't.
GitHub's documentation also says the Prevention tab on the overview dashboard shows AI Scan finding trends, resolved counts, and how often developers accepted a Copilot Autofix suggestion. The AI Scan tab covers findings over time, how often autofix suggestions were available, and the outcomes of findings with and without suggestions.
Availability and licensing
- The preview runs on github.com, including GitHub Enterprise Cloud.
- It is for customers with GitHub Code Security or GitHub Advanced Security.
- GitHub Enterprise Server is not supported in this release.
- GitHub's AI Scan documentation says that during the public preview, AI Scan requires a GitHub Advanced Security license and a GitHub Copilot license. Usage also consumes AI credits.
The changelog's availability line and the documentation's licensing line are worded differently. Confirm your own entitlements before planning around the feature.
What AI Scan covers
Understanding the scanner matters for reading the dashboard. The metrics only describe what AI Scan produces.
- Pull requests only. Unlike CodeQL alerts, AI Scan findings are only available on pull requests and do not appear as backlog alerts in the repository's security view. The new numbers are not a full inventory of a repository's vulnerabilities.
- Advisory, not blocking. AI Scan findings are advisory and do not block pull request merges. Per the documentation, they also can't yet be used in rulesets to enforce merge requirements.
- Independent of CodeQL. AI Scan does not require CodeQL default setup and does not depend on the state of CodeQL analysis, so it still runs if CodeQL fails or is waiting.
- Gaps CodeQL doesn't fill. GitHub points to languages and frameworks such as PHP, Shell/Bash, Terraform (HCL), Dockerfiles, JSP for Java and Blazor for C#.
- Exclusions. It does not run on pull requests from forks or on pull requests created by Dependabot.
- False positives. GitHub warns that, as with any AI-based tool, findings may include false positives. Each finding has thumbs up/down feedback.
The documented detection categories are:
- string injection
- weak cryptography
- broken access control
- sensitive data exposure
- security misconfiguration
- authentication failures
- data integrity failures
- server-side request forgery
- supply chain risks, such as unpinned third-party actions, packages or images
How it gets turned on
The metrics only mean something if the feature is enabled. GitHub's documentation says AI Scan is not allowed at the enterprise level by default and is disabled at the organization and repository levels. The documented order is:
- An enterprise owner allows AI Scan through the AI Scan policy under "Code Security".
- An organization admin turns it on under "Code scanning". This applies to eligible repositories where code scanning is already enabled.
- Repository admins can opt out of individual organization-owned repositories.
The documentation also says a REST API can manage the organization or repository ai-scan setting.
How this fits with the October 6 update
A week earlier, GitHub added AI Scan enablement status to the Coverage view. Repository rows show each repository's effective AI Scan enablement, and the view can be filtered with code-scanning-ai-scan-pr-scan:enabled and code-scanning-ai-scan-pr-scan:not-enabled. Coverage CSV exports also gained a Code Scanning AI Scan for pull requests column.
Put together, the two updates give a basic loop:
- Coverage view: where is AI Scan on?
- Code scanning on pull requests: what is it finding, and what are developers doing about it?
One caution from the documentation: the "not enabled" bucket can include repositories that are simply ineligible, and it doesn't say which reason applies. Don't read that count as a list of repositories someone forgot to switch on.
How to read the numbers
GitHub says the metrics help security teams evaluate AI Scan's effectiveness and support rollout and investment decisions. That is a reasonable use, with some limits. Here is what I'd keep in mind. This is my own analysis, not GitHub's guidance.
- Resolved isn't the same as fixed by AI. The changelog doesn't define "remediation rate" or explain how a resolution is attributed. Treat the figure as a trend indicator, not a precise measure, until GitHub documents the formula.
- Autofix availability isn't autofix adoption. The AI Scan tab separates findings that come with a suggestion from those that don't. GitHub's documentation says most findings include a suggested remediation, but not every one does. Applied suggestions are tracked separately.
- High counts can mean two things. A repository with many findings could be risky, or it could be a heavily active repository in a language the scanner covers. Compare findings against pull-request volume and language mix before drawing conclusions.
- False positives affect the data. If the scanner is noisy, "unresolved" findings may reflect developer skepticism instead of neglect. The feedback buttons are the documented way to flag inaccurate results.
- Unknowns. The announcement doesn't say whether these metrics can be exported or how configurable the reporting period is beyond the selection shown. Check the live interface.
Why admins should care
Rolling out an AI-based scanner across hundreds of repositories raises a practical question. Is it producing signal or noise, and are developers paying attention? Before this update, answering that meant sampling pull requests by hand. Now there is an aggregate view, split cleanly from CodeQL reporting, which keeps the deterministic and probabilistic engines from being blended into one number.
The caveats are real. This is a preview, it runs on pull requests only, it doesn't enforce anything, it doesn't support Enterprise Server, and it consumes AI credits. For teams that already use Advanced Security and Copilot and have PHP, shell, Terraform or Docker code that CodeQL skips, the dashboard is a sensible way to run a pilot. Enable it on a few repositories first, watch the AI Scan tab for a few weeks, and decide from there.