React2Shell, Agent Sprawl, and Your Vibe-Coded App

A critical React vulnerability just got a name, a CVE number, and an automated patch pipeline. At the same time, the coding agents behind Lovable, Cursor, Bolt, and similar tools are getting more autonomous, and founders who started with Supabase are quietly rebuilding their auth. None of this is a single fire alarm. Together it's a pattern worth five minutes of your attention.

TL;DR

• React2Shell (CVE-2025-55182) is a critical vulnerability in React that also affects Next.js and other frameworks built on it — Vercel is shipping automated patch PRs for it, per its security bulletin.
• Cursor shipped an SDK for building programmatic agents, self-hosted cloud agents, and "automations" — always-on agents that act on triggers without a human present.
• Vibe Coding Weekly #42 reports that 79% of agent-written pull requests are reviewed by the same person who wrote the prompt — meaning most agent code gets zero independent review.
• Developers are migrating away from bolted-on Supabase auth toward Clerk and now Better Auth, according to a widely discussed Val Town writeup — a sign that hand-rolled auth logic in AI-built apps keeps causing pain.

React2Shell: the CVE to check today

CVE-2025-55182, nicknamed React2Shell, is a critical vulnerability in React itself. Vercel's bulletin says it affects Next.js and other frameworks that sit on top of React — which covers a large share of apps generated by v0, Lovable, Bolt, and similar builders, since Next.js is a common default output. Vercel has also rolled out a unified security actions dashboard that flags affected projects, and automated patching that opens pull requests to upgrade vulnerable React and Next.js packages. If your app is hosted on Vercel, check the security dashboard. If it's hosted elsewhere, check your React and Next.js versions directly against the patched release named in the bulletin.

Agents that ship code with no one watching

Cursor's recent changelog reads like an autonomy escalation: an SDK so anyone can build agents on Cursor's own runtime, self-hosted cloud agents that run inside your own network, and automations — agents that trigger and act without a person kicking things off. That's more code moving with less friction. The problem is what happens to review once the agent both writes the code and merges it. Vibe Coding Weekly puts a number on it: 79% of agent-written pull requests are reviewed by the same person who wrote the prompt. That's not a second set of eyes. That's the same person checking their own homework, twice, quickly.

The quiet exit from Supabase auth

A widely shared Val Town blog post traces one team's path from Supabase auth to Clerk to Better Auth. The point isn't that Supabase is insecure — its anon key and default setup are public by design, the same way a Firebase web config is public by design. The real risk was never the key. It's what gets built around it: access rules, session handling, and row-level security policies that a founder wrote once, under time pressure, and never revisited. Teams making this move are choosing managed auth precisely because writing and maintaining that logic by hand is where mistakes creep in. If you're running Supabase, the question isn't whether your anon key is exposed. It's whether your row-level security policies actually restrict what that key can touch.

FAQ

Is my Next.js app affected by the React2Shell CVE (CVE-2025-55182)?

If your app was built with Next.js — a common default for apps generated by AI builders — check your React and Next.js versions against the patched releases named in Vercel's bulletin. Vercel-hosted projects can check the unified security actions dashboard for an automatic flag and a pre-built fix PR.

Should AI agents rewrite code without human review?

No, not without someone other than the prompter checking the diff. Vibe Coding Weekly found 79% of agent-written PRs are reviewed by the same person who prompted them, which isn't independent review. As Cursor and others ship more autonomous, always-on agents, this gap matters more, not less.

Why are founders leaving Supabase auth for Clerk or Better Auth?

Not because Supabase's public keys are a problem — they're public by design. Per the discussion around this Val Town post, teams move to managed auth providers because hand-written row-level security rules and session logic are easy to get wrong and easy to forget to revisit as an app grows.

The bottom line

Today's changes point the same direction: more automated defense on the infrastructure side (Vercel's patching and dashboard), and more autonomy with less review on the agent side (Cursor's SDK, automations, and the 79% same-person-review stat). If you ship with an AI builder, do two things this week: confirm your React/Next.js version isn't on the vulnerable list, and read your own row-level security or auth rules as if you didn't write them.