Vercel's new AI security guard checks every push, even force pushes

On November 2, 2026, Vercel opened Security Review to the public as a beta, a test version that is still being finished. It is an AI checker that reads new code and looks for security holes, such as a page that forgets to check who is asking, or a password left sitting in the code. It is part of Vercel Agent, the AI helper built into Vercel's dashboard. A day later, the changelog entry set out who gets it: teams on Vercel's Pro and Enterprise plans.

The detail that matters most is when it runs. Most code checkers wait for a pull request, which is a proposed change that someone reviews before it joins the main code. Security Review starts on every push instead. A push is the moment code is sent up to GitHub. Vercel's documentation says this means it also covers direct pushes and force pushes to the production branch, the live version of a site. A force push overwrites history, and it is the kind of change that can skip a normal review.

That matters for Jon. His site deploys on Vercel, and several of his agents send code to the site's repository. A checker that only watches pull requests would miss a change that goes straight in. Jon's Vercel team is on the Pro plan, so the beta is open to it. The fleet has not switched it on yet, so there are no results of its own to report.

Each review normally reads only the change, compared with a base. For a deployment, the base is the branch's last deployment. When there is no base, as on a project's very first review, it reads the whole codebase, which can cost more. Findings land in a Vulnerabilities dashboard. On private repositories they are also posted as comments on the pull request, and the docs warn that anyone with access to that repository can read them, code snippets included. That can be more people than can see the dashboard. On public repositories, the details stay in the dashboard.

How to use it

1. In Vercel, open the team's Agent settings and find the Security Reviews section.
2. Turn on Review Deployments Automatically, then choose all projects (the default) or up to 100 specific ones, and save.
3. To let Vercel Agent fix what it finds, also turn on Enable Chat on the same page. A fix opens a pull request, and nothing merges until a person approves it.
4. To hand findings to another coding agent, run vercel security vulnerabilities with Vercel CLI 65.10.0 or later.

Cost: reviews are free until November 10, 2026. After that, and for fixes right now, Vercel charges the AI provider's own price for inference tokens (the chunks of text a model reads and writes), plus $0.25 per million tokens, according to Vercel Agent pricing. Because it runs on every push, the bill grows with how often code is pushed.

The fine print: before turning it on for a private repository, check who can read it, because findings with code snippets will show up there.

What nobody knows yet

How well it finds real problems. Vercel's own example is a made-up invoice page, and no outside test has measured how many holes it catches or how often it raises false alarms. Vercel also says it reads code, not the running site, and does not promise a site is free of holes.

What to watch

November 10, 2026, when free reviews end. Then the first real bills, which will show what a busy repository costs per week.