Vercel's AI Security Reviewer Flags 6,029 Serious Flaws in Its First Week

In its first 24 hours, Vercel Security Review flagged 6,817 vulnerabilities in customers' code. Less than a week after its November 3 launch, the count had reached 57,897, including 6,029 rated critical or high, and Vercel's agent had already patched 14,665 of them. Guillermo Rauch, Vercel's chief executive, posted the numbers on November 10, the day the free trial ended. For the developers behind those repositories, thousands of holes were closed with a click and a review.

One Finding Every Ten Seconds

Vercel Security Review is in public beta for Pro and Enterprise teams. Once a team switches it on, every push to a connected GitHub repository is checked for vulnerabilities, including direct pushes to the production branch. Findings land in a Security dashboard, as a GitHub check on the commit and as comments on the pull request.

The first week's totals work out to roughly one finding every ten seconds, day and night. Rauch's first update, 24 hours in, counted 580 critical or high findings; the second counted more than ten times as many.

“I'm shocked at the number of critical vulnerabilities that would go unnoticed.”— Guillermo Rauch

The figures are Vercel's own, without a breakdown by type, and every severity rating is generated by the AI. The more telling question is what kind of flaw it finds, and how.

How It Catches What Pattern Scanners Miss

Each review covers only what changed, comparing the push against a base, which for deployments is the branch's previous deployment. That is why a review finishes in seconds and can run on every preview and production deployment.

The agent reads the diff with the code around it, traces how the changed code can be reached, checks for protections already in place and rechecks each finding before reporting it. Vercel says that is how it catches logic flaws such as missing authorization checks, which pattern-matching scanners miss. Its own example is an invoice download endpoint that never checks which organization an invoice belongs to, so any signed-in user could fetch another company's invoices by changing an ID.

Fixing takes one click. The agent confirms the vulnerability still exists, makes the smallest change that fixes it, updates the tests and opens a pull request, and nothing merges until a developer approves it. Teams that prefer their own coding agent can pull findings with the command vercel security vulnerabilities, and they can write security policies in plain language, such as requiring any endpoint that returns customer data to check that the requester owns it.

That last part matters more every month, for a reason Vercel spells out in its launch post.

Why Every Push Now Gets a Second Reader

Vercel's argument is that both sides of software security are speeding up. Frontier AI models make vulnerabilities easier to find and exploit, it notes: last month researchers turned two decade-old flaws in PHP's core into working remote code execution exploits. At the same time, agents are writing more of the code teams ship, and each handles security review its own way, or not at all.

Vercel is clear about the limits. The tool analyzes source code; it is not a security audit, it does no runtime or penetration testing and it cannot guarantee an application is free of flaws. Since November 10 it is no longer free: reviews are billed at the underlying model's token rate with no markup, plus $0.25 per million tokens.

In his post, Rauch says it works on any repository, wherever the code is deployed, and the launch post lists deployments from Vercel's command line as a next step. About a quarter of everything it found in week one is already fixed.

Security review used to be a gate at the end of a project; now it reads every push, and the fix arrives with the warning.

What people are saying

Guillermo Rauch@rauchg

“Vercel Security Review, less than 1 week since launch: ✓ 𝟱𝟳,𝟴𝟵𝟳 vulnerabilities identified ✓ 𝟲,𝟬𝟮𝟵 critical or high ✓ 𝟭𝟰,𝟲𝟲𝟱 already patched by Agent I'm shocked at ① the number of critical vulnerabilities that would go unnoticed and ② the ease at which the”

550149Nov 10