Vercel AI Updates: November 4, 2026

1. Vercel Security Review Enters Public Beta, Reviewing Every GitHub Push With Vercel Agent

Vercel launched Security Review in public beta for Pro and Enterprise teams. It was announced on November 2, with a changelog entry following on November 3. Security Review is a Vercel Agent capability alongside Code Review and Investigation. It runs on GitHub push events rather than pull requests, so direct and force pushes to the production branch are also reviewed. Each review compares the push with a base (for deployments, the branch’s previous deployment) and scans only the diff, so it finishes in seconds. In monorepos it is scoped to the project’s root directory. The agent reads the diff with the surrounding code, traces how the changed code can be reached, checks for existing protections and rechecks each finding before reporting it. Vercel says this catches logic flaws, such as missing authorization checks, that pattern-matching scanners miss. Teams can add plain-language security policies for the whole team or for individual projects, also through vercel security policy. Each finding records the project, environment, branch and commit, with a summary, a suggested fix and an AI-generated severity rating. Findings are deduplicated and follow a change into production when its pull request merges. On private and internal repositories they appear as a GitHub check and as pull request comments; public repositories keep details in the dashboard. Each new finding also emits an audit log event that Enterprise teams can send to a SIEM. “Fix with Vercel Agent” confirms the issue still exists, makes the smallest fix, updates tests and opens a pull request that a person must approve. Teams that prefer another coding agent can pull findings with vercel security vulnerabilities (Vercel CLI 65.10.0 or later) or export them as Markdown or JSON. Vercel notes that the tool only analyzes source code and is not a security audit, runtime test or penetration test. Reviews are free until November 10, 2026. After that they bill at the Vercel Agent rate: provider token prices with no markup, plus $0.25 per million input, cached and output tokens, so cost grows with how often a team pushes and how much code changes. Vercel plans to extend Security Review to deployments from the Vercel CLI and v0, and to repositories that do not deploy on Vercel. Source Changelog

2. Mercor Routes 63 Billion Tokens Through AI Gateway in 10 Days, and Zapier Creates 3.3 Million Sandboxes in a Month

Vercel published two customer stories on November 3. Mercor, which supplies expert human data and automated grading to AI labs, runs open-model batch jobs with 24-hour deadlines. Direct integrations and other gateways could not supply enough tokens-per-minute capacity, so requests piled up in retry queues and the jobs ran past 12 hours. With AI Gateway’s extra capacity and failover routing between providers, Mercor says the jobs now finish in 2 to 3 hours, and in its first 10 days it routed 18 million requests and 63 billion tokens without hitting rate limits. Mercor still uses other gateways as well. Hundreds of Mercor engineers reach open models through per-user keys with expiration dates, token limits and spend limits. Mercor applies the same controls to its internal agents and scopes each agent’s system and action permissions separately. Vercel Passport handles sign-in for about 2,000 projects, many built by business teams, with access granted from identity provider groups in seconds. Zapier reports creating 3.3 million Vercel Sandboxes in one month to isolate the code execution behind its AI workloads. It also uses AI Gateway as a single endpoint for the model requests it makes on behalf of customers, and builds interactive partner prototypes with more than 10 views in v0 in under two hours. All of these figures are customer-reported in case studies Vercel published. Mercor Zapier

3. Vercel Labs’ agent-desktop 1.1 Reaches the macOS Dock and Menu Bar and Re-Identifies Elements Through Accessibility Addresses

Vercel Labs released v1.1.0 of agent-desktop, its Apache-2.0 computer-use CLI for AI agents (about 13,300 GitHub stars). On macOS, agents can now reach the Dock, menu bar status items and system panels, and element refs are re-identified through each element’s own accessibility address, while Linux and Windows refs resolve through a capture rule, making refs from an earlier snapshot more likely to keep pointing at the same control. The release also adds a type --interval-ms flag for typing at a steady pace, moves the cursor before the agent takes the interaction lease, and polishes the on-screen overlay for humans watching or recording a session (an insertion caret after text entry, image cursors on Windows, and an accent-colored cursor label pill), alongside Windows fixes for commands crashing at exit, double-scaled overlay coordinates and CR LF line breaks; it also removes the macOS command host. Source