SI Agent News
DevelopingClaudeOpenAIAgentsSecurity

Three free tools check AI-written code for leaks, broken pages and ugly design

A dev.to guide shows how to give Claude Code, Cursor and Codex a security scanner, a browser tester and a design audit, so their work gets checked before it ships.

92security rules in the vibe-shield scanner

The short version

Why it matters

AI coding agents write code faster than most people can review it, and a build that passes can still hide leaked keys or a broken page. Automated checks are a cheap safety net, especially for non-engineers who can't read the code themselves. The setup is still one author's work, so start small and see what it catches.

The gap after the agent finishes

As dev.to reported, AI coding agents such as Claude Code, Cursor, OpenAI Codex, Windsurf and Cline can put together a full web application in minutes. The slow part is now checking it. The guide's author says that step is still mostly done by hand.

The post names three kinds of problems that slip through. The first is security: a secret key exposed to the public part of a site, or a payment notification accepted without proof it came from the sender. The second is that the code builds cleanly while the live page crashes, returns a server error or spills sideways on a phone screen. The third is design: washed-out contrast, uniform purple gradients and uneven spacing.

The author built three tools, one per problem. Each runs locally with a single command and comes with plugin files for Claude, Cursor and Codex. That lets the coding agent run the checks on its own work before each commit, meaning each saved set of changes.

vibe-shield: a security checklist

The first tool, vibe-shield, reads a project's files and applies 92 rules. The author says a full scan of a repository takes under a second. Nothing runs the app; it only reads the code, which is why this kind of check is called static.

The post groups the rules into eight areas and describes several in detail. One area is secrets that end up in code sent to the visitor's browser, such as a database service key or a payment secret. Another is webhooks, the automatic messages one service sends to another, which should be verified with a cryptographic signature. The tool also flags database tables with Row-Level Security turned off (a setting that limits which rows each user can see) and risky ways of building database queries.

It also covers AI-specific risks, such as prompt injection, where hidden text tricks a model into obeying the wrong instructions. A second command audits the instruction files that agents read, including SKILL.md, CLAUDE.md, .cursorrules and .mcp.json. It looks for hidden characters, encoded shell commands and attempts to send secrets out.

ghost-qa: a tester that uses a real browser

The second tool points at a local development server or a staging site and tests it with Playwright, software that drives a real web browser automatically. It looks for what a person clicking around would find, but a code review would not.

The author describes eight inspection traps:

The idea is that a build can succeed while the page itself is broken. Running a real browser is the only way to see that.

unslop: a design audit for AI-made interfaces

The third tool, unslop, applies 72 rules across seven design areas. It targets patterns the author sees as typical of AI-generated interfaces. It works with React, Next.js, Vue, Svelte and Tailwind CSS projects.

On colour, it checks contrast against the WCAG 2.1 accessibility standard, using 4.5:1 for AA and 7:1 for AAA, plus a newer method called APCA. It also flags pure black dark-mode backgrounds and overused indigo and purple gradients. On type, it looks for missing size scales, text lines wider than 75ch, and missing tabular-nums, a font setting that keeps digits aligned in tables. On layout, it flags overuse of Bento grids, nested scroll areas and inconsistent spacing.

It can also export ready-made style settings for ten design presets, such as linear-precision and swiss-grid-editorial.

How to read this guide

Each tool has an init command that sets up its rules and connection settings in a project. The connection uses MCP, the Model Context Protocol, a standard way to plug tools into AI assistants. The post ends by asking readers to star the projects on GitHub.

Keep in mind that the author built these tools and we haven't verified the rules or tested the setup. The post doesn't say how many false alarms the scanners produce, or how well the agents act on the results.

A sensible way in is to try one tool on one project and read what it reports. Aim it first at the pages that carry the most risk, such as login and payment, before you switch everything on.

What to watch

Try it

In a project you can afford to experiment on, run the author's scan command, npx -y @vibefoundry/vibe-shield@latest scan . , and read every finding yourself before trusting any of it.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by dev.to · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news