AI-Written Apps Are Shipping With Hundreds of Security Holes, Study Finds

Key points

• Researchers found 434 exploitable security flaws across a set of applications built using AI coding assistants.
• The most common problems included denial-of-service vulnerabilities (where an attacker can crash or overload an app), broken authorisation controls (where users can access things they should not be allowed to see), and secrets exposure (where passwords or private keys are accidentally left visible in the code).
• The findings suggest AI coding tools reproduce insecure patterns at scale, not just occasionally.
• No single vendor's tool was named as uniquely responsible; the problem spans the category.

Software is getting built faster than ever. AI coding assistants, sometimes called "vibe coding" tools because developers describe what they want in plain language and the AI writes the actual code, have made it possible for almost anyone to ship a working application in an afternoon. Speed comes with a catch.

A new analysis, covered by SecurityWeek, examined apps produced with these tools and found 434 flaws a real attacker could actually use. Not theoretical weaknesses. Practical ones.

The three biggest categories tell a clear story. Denial-of-service weaknesses let an attacker send a flood of requests that overwhelms an app and knocks it offline. Broken authorisation means the app fails to check who is allowed to do what, so one user might read another's private data or perform actions reserved for administrators. Secrets exposure is where the AI bakes passwords, private encryption keys, or database credentials directly into the code, where they can be found and stolen. We first looked at how AI tools leak credentials at scale on 10 July, and that reporting holds up against these numbers.

Why should ordinary users care about this?

If an app you use was built with an AI coding tool and holds your payment details or health records, these flaws create real doors for criminals. Broken authorisation alone could let a stranger read your private account information without ever guessing your password.

The problem is structural. AI models learn from enormous collections of existing code, and a lot of that code is insecure. The model reproduces those patterns confidently and at speed. A developer who is not a security expert, which describes many people now building apps with these tools, may have no way to spot the trouble before they hit publish.

From a threat-intelligence perspective, this matters beyond individual apps. At medium confidence, organised criminal groups and state-linked actors alike scan for exactly the flaw categories listed here. Secrets exposure is a well-documented initial-access technique: automated tools scrape publicly visible code repositories for leaked credentials, then use them to reach backend systems.

Should you worry?

If you use apps built by small teams or solo developers, some practical steps reduce your exposure. Watch for unexpected emails asking you to reset passwords you did not change. Check whether the app offers two-factor authentication, which requires a second code alongside your password, and turn it on. If the service holds financial data, review statements regularly.

Fast shipping is not the problem on its own. Shipping without a security review at the end is.