On November 2–3, 2026, Vercel announced Security Review. It is a new capability inside Vercel Agent, the company's AI agent product that already does Code Review and Investigation. Once enabled, it scans the changed code in every push to a connected GitHub repository for vulnerabilities such as missing authorization checks and exposed secrets. Findings show up in a dashboard, as PR comments, as GitHub checks and in the CLI. Vercel Agent can also propose fixes as pull requests.
This matters for builders because the security scan moves into the deploy platform and is billed per token. Teams shipping AI-generated code at high speed get a check on every push, including direct pushes to production. The open questions are coverage, accuracy and cost once the free period ends.
What we checked
- What concrete evidence would resolve the remaining editorial uncertainty for this item? partly confirmed. Vercel's blog post and changelog settle the factual questions: the product exists, it is in public beta for Pro and Enterprise, it is billed through Vercel Agent, it is free until November 10, 2026, and Vercel states its limits. They do not settle whether the scanner is any good. I found no independent benchmark, false-positive data or third-party testing. [1][2]
- Is there independent evidence of Security Review's detection accuracy or false-positive rate? unresolved. Searches turned up only Vercel's own pages and secondary rewrites of them. Vercel publishes no precision or recall figures, and it describes severity ratings as AI-generated. [2]
What we found
- Each review covers only the diff against a base. For deployments, the base is the branch's previous deployment. Vercel says this keeps reviews to seconds. The agent traces whether the changed code can be reached and rechecks each finding before reporting it. Reviews run on push events, not only on pull requests, so direct pushes and force pushes to production are covered. [1][2]
- Vercel describes the limits itself. Security Review analyzes source code only. It is not an audit, and it does not do runtime or penetration testing. Severity ratings are AI-generated. Because it reviews diffs, it is unclear whether it catches vulnerabilities that only appear when unchanged code interacts. [2]
- Reviews are free until November 10, 2026. After that, the blog post says they are billed at the Vercel Agent rate: model tokens at the provider's rate with no markup, plus $0.25 per million tokens. Fixes made with Vercel Agent are billed the same way. Total cost depends on how many pushes a team makes and how large each diff is, and Vercel publishes no per-review estimate. [2][1]
What it means
If you deploy on Vercel from GitHub, the free window until November 10 is a cheap way to measure Security Review on your own repos. Seed known flaws, such as a missing auth check or a leaked secret, and record the false-positive rate. Also estimate token use per push before billing starts. Treat it as an extra layer on every push, which matters most for code written or pushed by agents. It does not replace dependency scanning, runtime and DAST testing, or human review of auth logic. Vercel says as much.
The plain-language policies and the CLI/JSON export are the parts most useful for integration. You can encode house rules once and route findings to your own coding agent or SIEM. Before you make the GitHub check a merge gate, keep in mind that severity is AI-generated and that diff-only scope may miss issues that span unchanged code. No independent accuracy data exists yet.
Competing readings
- This is a real, incremental security layer built into the platform. An AI scanner that reviews every push and is tied to the deploy pipeline is useful for teams hosted on Vercel, but it adds to dedicated security tooling rather than replacing it (likely, about 60%). What would tell it apart: Every push is reviewed, including direct and force pushes to production. Before reporting a finding, the agent checks reachability and existing protections and rechecks the finding [1][2]; There are integration hooks: a CLI command (Vercel CLI 65.10.0 or later) and SIEM streaming for Enterprise teams [2][1].
- This is mainly a way to drive Vercel Agent token consumption. Its real security value is unproven because severity is AI-generated and no independent accuracy data exists (unlikely, about 30%). What would tell it apart: Reviews are free only until November 10, 2026. Fixes are billed at the Vercel Agent rate, and cost grows with push frequency [2][1]; Severity ratings are AI-generated, and diff-only scanning may miss issues that span unchanged code [2][1].
- This will displace dedicated AppSec and SAST vendors for Vercel customers (very unlikely, about 10%). What would tell it apart: Vercel says it is not a security audit and does no runtime or penetration testing. It scans only the diff, so it may miss issues spread across unchanged code [2][1].
Base rate
The New Runtime archive holds only 4 earlier announcement stories about companies or products from the last 12 months that are at least 3 months old: too few (minimum 5) for a base rate. The archive holds only 4 comparable announcement cases, and 3 of them were followed by a release or access change. That is below the minimum of 5, so there are too few cases to lean on for this story.
What would change this view
- Independent tests show high false-positive rates or miss common vulnerabilities such as missing authorization or leaked secrets (signal): It would weaken the useful-layer reading and support the reading that this is mainly a monetization play.
- Vercel extends the free period or changes Agent token rates after November 10, 2026 (pricing signal): Adoption is cost-sensitive, and Vercel is still positioning the product rather than monetizing it.
- Support ships for repositories not deployed on Vercel, or for GitLab and Bitbucket (signal): Vercel is moving to compete with standalone AppSec tools, which would raise the odds of the displacement reading.
Still open
- What are the precision and recall on standard vulnerability benchmarks? Neither Vercel nor anyone independent has published figures.
- Which models power the reviews, and does code leave Vercel's infrastructure for third-party model providers?
- What does a typical review cost in tokens once billing starts on November 10, 2026?
- Does diff-only scanning miss vulnerabilities created when changed code interacts with unchanged code?