GitHub makes agentic autofix and AI Scan generally available for code scanning

What GitHub made generally available on 15 October

GitHub published two related changelog posts on 15 October 2026: Agentic autofix for code scanning alerts general availability and AI Scan for pull requests generally available. Together they push GitHub’s security agents from preview language into production packaging for customers already paying for Code Security or Advanced Security on Enterprise Cloud.

Agentic autofix is the remediation path. GitHub’s description is that it works across the codebase the way a developer would: explore relevant files, propose a fix, re-run CodeQL to confirm the alert is closed, iterate if needed, then open a draft pull request for human review. The pull request is supposed to include a summary of the fix, why it closes the alert, and the validation steps Copilot took. Reviewers can keep steering the same session from pull request comments or the repository’s Agents tab.

AI Scan is the detection path for pull requests. It sits beside CodeQL rather than replacing it, targeting languages and frameworks CodeQL does not currently cover. The changelog’s examples include PHP, Shell/Bash, Terraform (HCL), Dockerfiles, JSP for Java, Blazor for C#, and more. That pairs with other recent GitHub agent security moves we have covered, including HydraFusion Max multi-model orchestration for Copilot and third-party agentic AppSec such as Snyk Evo’s security agents.

How agentic autofix differs from Generate Fix

GitHub’s autofix documentation draws a bright line between agentic autofix and Copilot Autofix (the older Generate Fix experience). Both use models plus codebase context and code scanning analysis. They differ in workflow, results, access and billing.

When agentic autofix is enabled, it replaces the free Generate Fix option on individual alerts with an Assign to Copilot button. You can assign a single alert on the repository’s default branch, select up to 25 alerts from a repository or security-campaign list for one combined pull request, or call the Update a Code Scanning Alert REST API with assignees set to ["copilot-swe-agent[bot]"]. Alerts that exist only on pull requests or non-default branches cannot be assigned.

Copilot Autofix remains the lighter path: suggested fixes that do not consume AI Credits, enabled by default wherever CodeQL code scanning is on unless an administrator disables it, including public repositories on GitHub.com. Agentic autofix is the end-to-end agent session that explores beyond the affected file, validates, and files a draft PR—at the cost of AI Credits and Actions minutes.

Who can turn it on, and what it costs

Access is layered. The changelog says an enterprise administrator must enable the new “Agentic autofix” policy under Advanced Security policies. After that, an organization or repository administrator turns the feature on in Code Security settings. A Copilot billing method must also be defined. Docs add nuance for public repositories owned by personal accounts, where a repository administrator can enable agentic autofix directly, and note that repositories owned by Enterprise Managed Users are out of scope for agentic autofix even though Copilot Autofix can still apply there.

Billing is explicit. Agentic autofix consumes AI Credits from the Code Security budget and GitHub Actions minutes. Usage shows under the AI usage section of enterprise billing as the Advanced Security product, and administrators can set SKU-level budgets for those credits. AI Scan likewise consumes Code Security AI Credits. Neither changelog prices a single fix in dollars; the meter is credits and Actions time.

That packaging matters for security teams already drowning in model-found findings. Bank remediation queues and vendor scanners have shown that discovery scales faster than patching. GitHub’s bet is that an agent which closes the loop—explore, patch, re-query CodeQL, open a PR—reduces human coordination cost, while still requiring a reviewer before merge.

What AI Scan adds on pull requests

AI Scan runs automatically on pull requests when code scanning and the repository’s effective AI Scan setting are enabled. Findings appear on the Conversation and Files changed tabs next to CodeQL results, with an AI indicator so reviewers can tell them apart. A finding can include an explanation of the security risk and a suggested autofix.

Availability is broader than agentic autofix on one axis and narrower on another. Public repositories owned by organizations or personal accounts can use AI Scan; private and internal organization-owned repositories need GitHub Code Security or Advanced Security. A Copilot license is not required for AI Scan. It does not run on pull requests from forks or on Dependabot pull requests.

Most important for governance: AI Scan findings are advisory. They do not fail the code scanning check and cannot block merges through ruleset requirements. Teams that want a hard gate still need CodeQL or other blocking checks; AI Scan is coverage expansion, not a new merge veto.

Where the two features fit a security workflow

A practical split emerges. Use CodeQL for languages it supports and for blocking policy. Use AI Scan to catch PR diffs in PHP, shell, Terraform, Dockerfiles and similar gaps. Use agentic autofix when you want Copilot to own a default-branch backlog item through a validated draft PR. Keep Copilot Autofix for cheap suggestions when you have not enabled the agentic policy. That stack sits alongside containment-minded tools such as Chainguard Harness’s Mythos-powered vulnerability hunting, which attacks the discovery side from a different vendor angle.

Security campaigns are a first-class entry point: assigning up to 25 alerts into one pull request is aimed at backlog burns, not at drive-by single-line typos. The REST assignee hook lets automation start sessions without clicking through the UI, which is how large orgs will actually wire this into existing triage bots.

Limits GitHub states up front

GitHub documents several hard limits. Agentic autofix works on a best-effort basis. Copilot validates fixes by re-running CodeQL with the code scanning query suite; it cannot confirm fixes for alerts from custom CodeQL queries, the security-extended query suite, or third-party code scanning tools. Default-branch-only assignment means PR-only findings stay outside the agentic loop until they land on default.

AI Scan’s advisory status is both a safety valve and a coverage gap: you get signal on languages CodeQL misses, but you cannot make that signal a required check today. Forks and Dependabot PRs are excluded, which reduces some supply-chain noise and also leaves those change paths unscanned by this feature.

What to verify before you rely on it

Before treating either GA label as a staffing plan, verify three things on your tenant: the enterprise Agentic autofix policy is on; org/repo Code Security settings match your intent; and Code Security AI Credit budgets will not silently stop sessions mid-campaign. Then sample a handful of draft PRs: did CodeQL actually clear the alert, did the patch match team style, and how often did humans reject or rewrite the change?

Independent measurements still missing from the changelog: median time from assignment to draft PR, false-fix fix rate after human review, Actions minute cost per closed alert, and AI Scan precision on Terraform or PHP compared with specialist scanners. Until those numbers exist outside GitHub’s marketing, the honest story on 15 October is product availability—agentic remediation with CodeQL validation, plus advisory PR scanning for CodeQL’s blind spots—not a proven reduction in production incidents.