Back to Blog

Vibe Coding Has a Security Problem, and the Numbers Are Finally In

• 3 min read • Ahmed Daoudi

For two years the pitch for vibe coding was speed: describe the app, let the agent write it, ship by the afternoon. In 2026 the security research caught up with the hype, and it does not look good for anyone shipping straight from the prompt to production.

What the research actually found

A team publishing on arXiv collected 9,041 open-source applications built with popular AI agents, Claude Code and Lovable among them, then manually audited 200 apps that were publicly deployed. They reported 1,186 vulnerabilities in that sample, which is roughly six per live app.

The Cloud Security Alliance reached a similar conclusion from a different direction. Its research notes put the failure rate of AI-generated code samples on security tests somewhere between 45% and 70%, depending on the method. The most common failures were broken authorization, missing access controls and hardcoded credentials. The same notes cite Georgia Tech's Vibe Security Radar, which counted CVEs directly attributable to AI coding tools rising from 6 in January 2026 to 35 in March 2026, and the researchers think the real number is several times higher because most tools leave no trace in commit history.

None of these bugs are new

Read the vulnerability lists and you will not find exotic attacks. You will find the OWASP Top 10 from a decade ago: an API route that checks whether you are logged in but not whether the record belongs to you, a database policy left wide open, an API key pasted into a frontend bundle.

The difference is the process around the code. When a junior developer writes an insecure endpoint, someone reviews the pull request. When an agent writes forty files in four minutes and the person driving it is a founder who cannot read the code, there is no reviewer. The model is fast, and the human in the loop has quietly turned into a human watching the loop.

A practical checklist if you ship AI-written code

My take

I use AI agents every day, and they make me faster. But the speed is only real if the review stays. On the HR and payroll systems I work on, one authorization bug means one employee seeing another's salary, and no deadline is worth that. Vibe coding is fine for prototypes. For anything holding real user data, the agent writes the draft and an engineer owns the result.

Sources