GitHub Agentic Autofix GA: CodeQL Fixes, Draft PRs, AI Credits and Enterprise Setup

Instead of merely offering a static snippet and wishing a developer good luck, this release turns vulnerability remediation into an autonomous, multi-step workflow. By handing off backlog alerts on a repository's default branch to GitHub Copilot, teams can let an AI agent traverse the repository, formulate a patch, test its own work against CodeQL, and package the result into a draft pull request ready for human inspection.

For enterprise administrators, DevSecOps engineers, and software leads, this general availability release also introduces a much stricter set of governance controls, branch boundaries, and usage-based billing rules than the earlier preview. Here is a complete breakdown of how agentic autofix works, how it compares to traditional Copilot Autofix, and what IT and security administrators need to configure before turning the agent loose on their repositories.

From Static Suggestions to Autonomous Sessions​

To appreciate what changes with this release, it helps to distinguish between the two distinct remediation experiences documented in GitHub Enterprise Cloud Docs: Copilot Autofix (the classic experience that reached general availability in August 2024) and the newly generally available agentic autofix.

Classic Copilot Autofix is a one-shot suggestion engine. When CodeQL spots a vulnerability, Copilot Autofix generates a single suggested code change and an explanation for that specific alert, which a human developer can review and turn into a pull request. It is enabled by default on repositories using CodeQL and does not consume AI Credits.

Agentic autofix, by contrast, operates like a junior developer assigned a ticket:

1. Codebase Exploration: Rather than looking exclusively at the flagged line of code, the agent starts a session and explores relevant files across the broader repository to understand dependencies, helper functions, and context.
2. Patch Generation: It drafts a proposed code change to remediate the vulnerability (taking into account any custom Copilot instructions configured for the repository).
3. CodeQL Validation and Iteration: Before bothering a human reviewer, the agent reruns CodeQL to verify that the proposed change actually resolves the underlying alert. If the check fails, the agent iterates on its work.
4. Draft Pull Request Creation: Once validated, Copilot opens a draft pull request containing the code diff, a plain-English summary of what was changed, an explanation of why the change closes the alert, and a log of the validation steps taken.

Once that draft pull request is open, the workflow remains interactive. If a reviewer notices that the agent violated an internal style convention or missed an edge case, the developer does not have to take over the branch manually right away. Reviewers can ask Copilot to refine the patch simply by leaving a comment on the pull request or interacting directly with the agent's session log inside the repository's Agents tab.

The Fine Print on Automated Validation​

Any seasoned security engineer knows that "the scanner stopped complaining" is not synonymous with "the application is secure and functional." While agentic autofix reruns CodeQL to confirm its fix closes the alert, GitHub Enterprise Cloud Docs explicitly notes that this validation operates on a best-effort basis.

Specifically, Copilot validates fixes using the standard code scanning CodeQL query suite. It cannot confirm fixes for alerts triggered by:

• Custom CodeQL queries.
• The security-extended CodeQL query suite.
• Third-party static analysis tools (such as SARIF uploads from external scanners).

If the agent cannot validate a fix, or if its exploration leads it to suspect the alert is actually a false positive, GitHub Docs notes that Copilot will explicitly state that caveat inside the pull request description. Human code review and full CI/CD regression testing remain mandatory before merging.

Three Ways to Assign Alerts (Including Bulk Campaigns)​

When agentic autofix is enabled on a repository, it replaces the free Generate fix button on individual code scanning alerts with an Assign to Copilot button. Users with write access to the repository can trigger the agent through three distinct methods:

• Single-Alert Assignment: Navigate to the repository's Security and quality tab (or Security view), select Code scanning, open an individual alert on the default branch, and click Assign to Copilot under the Assignees section.
• Bulk Backlog and Security Campaign Assignment: If you are tackling a cluster of related vulnerabilities, you can select up to 25 alerts at once from either the repository's security alerts list or a coordinated security campaign. Clicking Assign to Copilot launches a single agent session that attempts to remediate all selected alerts together within a single draft pull request.
• Automated REST API Assignment: DevSecOps teams building automated triage pipelines can trigger the agent programmatically via the Update a Code Scanning Alert REST API endpoint by setting the assignees array payload to ["copilot-swe-agent[bot]"].

The Default-Branch Boundary​

A critical architectural limitation highlighted in the October 15 GA announcement is branch scope: agentic autofix only supports code scanning alerts that exist on the repository's default branch.

If a code scanning check catches a brand-new vulnerability on a feature branch or inside an open pull request, you cannot assign that alert to the agentic autofix bot. The feature is engineered specifically to burn down established default-branch security debt, not to act as an inline fixer for unmerged feature branches.

Governance, Rollout Requirements, and Billing Changes​

If your organization tested agentic autofix during the July 2026 public preview, pay close attention: the administrative and billing plumbing has matured significantly for general availability.

According to the GitHub Changelog GA announcement, agentic autofix is available to GitHub Code Security and GitHub Advanced Security customers on GitHub Enterprise Cloud, including GitHub Enterprise Cloud with data residency. Notice what is not required at the individual seat level: according to GitHub Docs, individual developers assigning alerts do not need a personal GitHub Copilot license to use agentic autofix, though an enterprise Copilot billing method must be configured.

How to Enable Agentic Autofix​

Unlike classic Copilot Autofix—which is on by default when CodeQL is active—agentic autofix is gated behind a strict multi-tier opt-in hierarchy so enterprises do not accidentally rack up compute charges:

1. Define a Billing Method: Ensure a valid Copilot usage-based billing method is defined at the enterprise level.
2. Enable the Enterprise Policy Gate: An enterprise administrator must first unlock access by enabling the new "Agentic autofix" policy located under Advanced Security policies.
3. Activate at the Organization or Repository Level: Once the enterprise policy allows it, an organization owner, security manager, or repository administrator must turn on agentic autofix within the Code Security settings for the target organization or individual repository.

Watching the Meter: AI Credits and Actions Minutes​

Why the two-key administrative turn-on? Because every time someone clicks Assign to Copilot—or fires off a batch of 25 alerts via the REST API—two meters start running:

• Code Security AI Credits: Agent sessions draw down AI Credits from your organization's Code Security budget. In enterprise billing and licensing, this consumption is itemized under the "AI usage" section with "Advanced Security" listed as the product. To prevent surprise bills from overzealous backlog grooming, administrators can configure SKU-level budgets specifically for Code Security AI Credits.
• GitHub Actions Minutes: Because the agent spins up an environment to explore files and rerun CodeQL analysis, every remediation session also consumes GitHub Actions minutes.

Practical Takeaways for Engineering and Security Teams​

GitHub's move to make agentic autofix generally available represents a pragmatic evolution in application security. Generating a five-line patch for a SQL injection or cross-site scripting flaw was never the hardest part of remediation; the friction has always been checking whether that change breaks callers in three other files, running the scanner again to prove the fix works, and writing up the pull request.

To get the most value out of agentic autofix without blowing through your Actions minutes or AI Credit budgets, consider a phased adoption strategy:

• Enable selectively first: Use repository-level Code Security settings to enable agentic autofix on a handful of high-priority repositories with well-tested CI pipelines before flipping the switch across an entire GitHub organization.
• Set SKU-level budgets immediately: Before granting repository admins access, configure SKU-level budgets for Code Security AI Credits in enterprise billing so you have a hard ceiling on consumption.
• Group similar CWEs when batching: While the UI lets you bundle up to 25 alerts into a single pull request, stuffing 25 unrelated vulnerabilities into one PR creates a nightmare for human code reviewers. Use filters or security campaigns to batch small groups of identical vulnerability types (for example, five path-traversal alerts in the same module) so the resulting draft pull request stays cohesive and easy to audit.

Summary Checklist for Administrators​

• Availability: Generally available as of October 15, 2026, for GitHub Code Security and GitHub Advanced Security customers on GitHub Enterprise Cloud (including data residency deployments).
• Capabilities: Explores the codebase, proposes a fix, reruns standard CodeQL queries to verify closure, and opens a draft PR for up to 25 default-branch alerts at a time.
• Prerequisites: Requires the enterprise "Agentic autofix" Advanced Security policy enabled, org/repo Code Security toggle enabled, and a defined Copilot billing method.
• Cost Controls: Consumes Advanced Security AI Credits (manageable via SKU-level budgets) plus GitHub Actions minutes.