GitHub's November 13, 2026 update makes repository AI Scan settings manageable through its REST API. The operational opportunity is consistent rollout, not a guarantee that enabling a switch has examined your code. This guide separates permission, configuration, execution and verified remediation, then proposes a small rollout worksheet. Published November 15, 2026; no Nerova scan benchmark or customer result is claimed.
What the new AI Scan REST API actually changes
The public preview adds repository_scan to organization and repository AI Scan GET/PATCH endpoints. An organization can enable the setting for eligible repositories; repositories can opt out, but cannot opt in against a disabled organization setting. Enterprise policy can block enablement. When a combined pr_scan and repository_scan update is blocked, GitHub says neither setting changes. Enabling repository scanning does not start a scan. See the November 13 announcement.
For an operator, that means an infrastructure-as-code change should produce a configuration receipt, not a successful-security-scan badge. Keep those two records separate even if your internal dashboard eventually displays both.
Check eligibility, billing and analysis limits first
GitHub's current concept guide includes organization-owned public repositories and private/internal repositories with GitHub Code Security or GitHub Advanced Security. Personal public repositories are also eligible; personal private repositories are not. A Copilot license is not required, but scans consume Code Security AI credits. Personal-public billing can fall to an organization or enterprise that supplies the owner's Copilot license. Check the actual billing account, included credits and budgets.
Scans are manually triggered by users with write access, examine selected code on the default branch, and are neither exhaustive nor guaranteed accurate. Scheduled/on-push scans and self-hosted runners are unsupported. Detection steps have 30-minute and 150,000-credit limits; these are not a fixed whole-job price or completion promise. After merging a potential fix, manually rescan; alert closure can depend on changing the exact triggering line. Treat findings as review inputs, not proof of remediation.
A useful preflight question is therefore not just “can this repository use the feature?” Ask who can authorize the spend, who will receive the findings, and who can verify the deployed fix. If any answer is unknown, record that uncertainty rather than letting a bulk configuration run decide implicitly.
Use a read, plan, approve, verify rollout
The REST reference documents these settings endpoints:
GET /orgs/{org}/code-scanning/ai-scanPATCH /orgs/{org}/code-scanning/ai-scanGET /repos/{owner}/{repo}/code-scanning/ai-scanPATCH /repos/{owner}/{repo}/code-scanning/ai-scan
Organization reads require an owner or security manager; fine-grained tokens use organization Administration read permission for GET and write permission for PATCH. Repository operations have their own permission requirements. Check the specific endpoint rather than copying an organization token scope everywhere. A 422 response is not uniquely an enterprise-policy diagnosis: the reference also documents validation or abuse-related rejection.
We recommend a staged workflow rather than an unattended fleet-wide toggle:
- Read: collect the authorized current settings and inventory the intended repository set. Do not include unrelated repositories simply because a token can reach them.
- Plan: show the exact before/after values and identify which existing pull-request setting must remain unchanged.
- Approve: obtain the responsible owner's agreement to the settings change and the separate scanning budget. A technical ability to write a setting is not the same as organizational approval.
- Apply: make a small, bounded change. Retain response status and a correlation identifier without putting access tokens in logs.
- Verify: read the setting again. Record blocked and unknown outcomes separately; do not convert every non-success into a retry loop.
- Run and review: arrange an explicitly authorized scan, then assign triage and remediation work. This is a separate operational step.
A rollout ledger that does not confuse enabled with covered
Use distinct columns for repository identity, business owner, requested configuration, verified configuration, scan occurrence, source revision, findings reviewed, and remediation verification. Include an “unknown” state in each stage. A missing scan record should not be interpreted as zero findings, and zero findings should not be interpreted as a clean security audit.
For example, a hypothetical pilot might contain 12 approved repositories, 10 verified settings changes, seven completed scans and five completed triage reviews. Reporting “about 83% rolled out” describes configuration only: 10 divided by 12. Reporting “about 83% secured” would invent a conclusion those records do not establish. Label the denominator and the stage next to every percentage.
Keep deployed software in view as well. A repository's default branch can differ from the revision currently serving customers. Your own release inventory should identify that difference before an internal report implies the deployed artifact was assessed.
Measure whether the extra workflow is worth keeping
For a bounded pilot, track incremental reviewed findings, duplicate-review effort, false-positive disposition, time to owner assignment, credit consumption, and verified remediation outcomes. Do not reward a scanner merely for generating a larger backlog. Nor should a smaller backlog automatically count as improvement if scan coverage or reporting failed.
Choose the evaluation window before comparing runs. Record source revision, analysis configuration and triage criteria so a later result is interpretable. If you change the repository set and the review standard simultaneously, a simple before/after percentage will not identify which change mattered. Preserve separate cohorts or write down the confounding changes.
Extend the pilot only when its evidence supports the operating cost and the team can handle the resulting work. Our recommendation is an evaluation method, not a claim that this preview detects more vulnerabilities or saves money in your environment.
How this differs from Dependabot monitoring
Do not combine every security signal into one unlabeled health score. Our Dependabot December 8 readiness guide addresses a different monitoring-policy change. Keep each workflow's scope, owner and evidence visible. The common operational discipline is to distinguish a configured control, a completed observation, a reviewed finding and a verified fix.