New Tool Catches Hidden Bugs in AI-Coded Apps

Have you ever used an app that looks perfectly fine but just doesn't work? Maybe a payment button does nothing, or a form doesn't save your information. That's what researchers call a "silent failure." These problems are common in "vibe-coded" apps — software built by typing a description into an AI tool like Claude, so people who don't write code can create working websites. The catch is that AI isn't perfect, and the people using it often can't spot flaws in the code.

FlowCheck solves this by letting you describe what your app *should* do in plain English, directly through the interface you already use. For example: "When I click 'Add to Cart,' the cart count should go up." The tool turns that rule into an automated check that reads your app's code and verifies the rule is really true. No code reading required. You simply see the checks as familiar, human-readable statements on screen.

The researchers tested FlowCheck on four apps built with AI and injected 30 bugs. FlowCheck caught every single one with zero false positives — meaning it never wrongly flagged something as broken. In contrast, when they asked three leading AI models (Claude, DeepSeek, and Gemini) to find bugs in the same code, none achieved perfect accuracy. FlowCheck was more reliable than the very models used to write the code.

This isn't a product yet, but it points to a safer future for AI-generated software. Small businesses, teachers, and hobbyists who build tools with AI could rely on them without fear of hidden breakdowns. Instead of trusting that the AI "probably got it right," you'd have proof your app works as intended — protecting your time, your customers, and your reputation.