GitHub tags AI-generated code scanning alerts on the PR itself

The tab I kept open on the second screen

For years I have kept a code scanning tab pinned on a second monitor. It sat there because the alerts almost never showed up where I actually spent my day: in a security view, one link away from the PR I was reviewing, and one context switch too far for anyone in a hurry.

That is finally changing. GitHub's code scanning now surfaces AI-powered security detections directly on pull requests, alongside CodeQL findings. The alerts carry an AI label so you can tell them apart from the deterministic CodeQL ones, and they show up where reviewers already look.

What the changelog actually says

The July 14 GitHub changelog is short. Code scanning gains a new class of alert, generated with AI, that extends coverage to languages and frameworks CodeQL's built-in analysis does not support today. No specific languages are named in the announcement. If your stack is one of the ones CodeQL skipped over, you may finally see something appear on a PR you never saw a signal on before.

Two behaviours are worth pinning down before you touch a toggle. First, alerts appear with an AI tag next to CodeQL results, on the same PR surface, in the same review flow. Second, the changelog says plainly that these findings are informational and will not block pull request merges. That second sentence matters more than it reads on the page. I will come back to it.

What it takes to switch on

There is a two-step gate. AI security detections have to be allowed at the enterprise level first, and then enabled at the organization or repository level. Under the hood the repo also needs CodeQL default setup turned on, plus a GitHub Copilot license, since the AI runs draw down your organization's AI credits.

That is a lot of prerequisites for a checkbox, but the shape is familiar to anyone who has rolled out GitHub security features before: enterprise policy first, org policy second, repo enablement third. If you have already deployed CodeQL default setup, this drops in behind the same gate you already opened. If your enterprise has been holding the line on Copilot, this feature is not on its own a reason to reverse that decision, and the changelog is explicit that the credit spend is real.

How the label reshapes review

Here is the honest part. A non-blocking alert is a suggestion, and suggestions get ignored under pressure. My worry with any advisory finding on a PR is not that it fails to fire; it is that reviewers learn to scroll past the AI tag the way they have learned to scroll past a hundred other lint warnings.

Two things make me a little more hopeful. One, the labelling is explicit. You can see at a glance which alert came from a deterministic CodeQL query and which came from a model, which is the trust signal a reviewer needs before they spend a minute investigating. Two, keeping AI findings out of required checks is the right call today. Nobody wants a probabilistic security signal to be the reason a hotfix at 5pm cannot merge.

If your team already runs a triage rota, you will feel the difference on the first Monday after enablement. The AI-tagged alerts either turn into a real chore or they surface something CodeQL never caught. You will know inside a sprint.

Rough edges worth flagging

Cost is the first one. AI credits are a metered resource, and a busy repo will burn them. Watch which repos you enable this on, and whether the volume of findings justifies the spend.

Signal-to-noise is the second. CodeQL false positives are already a chore; a model-authored finding that a reviewer cannot easily reproduce locally will feel worse. Plan a couple of weeks to calibrate how you treat the AI-tagged alerts internally before you tell the wider org they can trust them.

Third, the informational-only status is today's floor, not a promise. At some point a team will want to promote a subset to required checks. GitHub does not describe a path to that in the changelog, and I would not design branch protection today on the assumption it will arrive.

What I want to see land next

Two things. Which languages actually start seeing coverage first, because that is where the real DX story lives for teams outside CodeQL's happy path. And how organizations end up wiring the AI-tagged alerts into their existing triage. If reviewers start acting on them within a sprint of enablement, this is a genuine win for the PR feedback loop. If the tag becomes background noise on the review page, the feature quietly joins the list of security dashboards nobody reads.