GitHub made Agentic Autofix for code-scanning alerts generally available on October 15. The feature can investigate relevant files, propose changes, rerun CodeQL and open a draft pull request, moving it beyond the one-shot patch suggestion offered by Copilot Autofix.
The important limit is in the word “validate.” GitHub says the agent works on a best-effort basis and cannot confirm every class of alert through its CodeQL rerun. The draft pull request is a review artifact, not a merge decision.
One assignment can cover 25 alerts
A maintainer can assign a code-scanning alert to Copilot from the repository’s default branch. GitHub also allows up to 25 alerts to be selected from an alert list or security campaign for one remediation pull request. The agent explores files beyond the line that raised the finding, iterates on a proposed change, and records a summary and validation steps in the resulting draft.
That broader context is the practical difference from Copilot Autofix. The older flow generates a suggested change for one supported alert and leaves the maintainer to create the pull request. Agentic Autofix runs a repository-level session intended to carry an alert or small alert group through investigation and validation.
The branch and alert source define the boundary
The general-availability release supports alerts on the default branch. Findings detected only on a pull request or another branch cannot be assigned to the agent. Teams that use short-lived release branches or PR-only scans should not assume the backlog has identical automation coverage.
GitHub’s documentation also draws a line around the validation rerun. Agentic Autofix reruns CodeQL with the standard code-scanning query suite. It cannot confirm fixes for custom CodeQL queries, the security-extended suite or alerts produced by third-party scanners. A patch may still be useful in those cases, but the relevant scanner or test must be run separately.
Enablement and cost sit outside the pull request
Agentic Autofix is available to GitHub Code Security and GitHub Advanced Security customers on GitHub Enterprise Cloud, including data-residency deployments. GitHub requires enterprise policy permission where applicable, organization or repository enablement, access to the Copilot cloud agent, and a configured Copilot billing method.
Each agent session consumes Code Security AI Credits and GitHub Actions minutes. Copilot Autofix suggestions do not consume AI Credits, so a team choosing between the two is choosing between a narrower suggestion and a longer-running remediation session with repository exploration and automated validation.
Review should target what the rerun does not prove
A cleared CodeQL alert shows that the selected standard query no longer reports the same condition in the analyzed code. It does not prove that authorization behavior is unchanged, that a new input path is safe, or that a third-party scanner agrees. GitHub explicitly tells maintainers to review and test generated changes before merging.
The highest-value review therefore follows the affected trust boundary: reproduce the original exploit condition when possible, inspect nearby callers and error handling, run the project’s tests, and rerun any scanner whose alert was outside the standard query suite. Agentic Autofix can compress the path from backlog item to reviewable patch, but the evidence required to merge still belongs to the repository’s normal security process.
