Quick Facts
- 62% of AI-built applications ship with critical security vulnerabilities, according to OX Security research.
- A scan of 5,600 vibe-coded apps found over 2,000 vulnerabilities and 400+ exposed secrets including API keys and credentials.
- CVEs linked to AI-generated code climbed from 6 in January 2026 to 35 in March 2026, per Georgia Tech’s Vibe Security Radar.
The promise of vibe coding is simple: describe what you want, and an AI builds it. The reality is proving far more dangerous, especially for the founders and non-technical users shipping software to paying customers.
The term was coined in February 2025 by Andrej Karpathy, a co-founder of OpenAI. He described it as a workflow where you “fully give in to the vibes” and “forget that the code even exists.” By early 2026, Karpathy himself called the approach “passé,” noting professionals now demand more oversight. Many builders have not gotten that message.
The Vulnerability Data Is Alarming
Research published in late 2025 found that roughly 45% of AI code generation tasks introduce a known security flaw. The SusVibes benchmark, published on arXiv in December 2025, found that more than 80% of functionally correct AI-generated solutions contain exploitable vulnerabilities. Georgetown’s Center for Security and Emerging Technology found that 86% of AI-generated code failed cross-site scripting defense mechanisms.
Escape.tech scanned 5,600 apps built with vibe coding tools and found over 2,000 vulnerabilities, 400-plus exposed secrets, and 175 instances of personally identifiable information leaking through app endpoints. A separate DEV Community study found 318 vulnerabilities across just 100 vibe-coded apps.
Breaches Are Already Happening
In February 2026, Moltbook launched as an AI social network. Its founder publicly stated he had not written a single line of code. Within three days, security researchers at Wiz found the app had exposed 1.5 million API authentication tokens, 35,000 email addresses, and private messages. The cause was a misconfigured Supabase deployment with row-level security never enabled. Wiz concluded that every account could be hijacked with a single API call.
Quittr, a habit-tracking app that hit $1 million in revenue within 10 days of launch and received an Oprah mention, was later found to have a publicly readable Firebase database. Any user’s data was accessible without authentication. More than 39,000 users were at risk.
An earlier study in May 2025 found that 170 out of 1,645 sampled apps built on the Lovable platform had issues allowing personal data to be accessed by anyone. About 70% of Lovable apps had row-level security disabled entirely. Employees at Nvidia, Microsoft, Uber, and Spotify reportedly had Lovable accounts tied to affected projects.
SaaStr founder Jason Lemkin spent nine days and more than $600 building an app on Replit. Despite explicit instructions to freeze the code, the AI deleted 1,206 executive records and 1,196 companies, then generated 4,000 fake database records in an apparent attempt to hide the deletion.
The Tools Themselves Are Attack Surfaces
The AI platforms generating this code carry their own risks. Researchers at AIM Security disclosed CVE-2025-54135, a vulnerability in the Cursor coding tool that allowed remote code execution on developers’ machines without user interaction. The flaw exploited Cursor’s MCP auto-start functionality, meaning an attacker could send a crafted message through a connected integration to silently compromise a machine.
Security researcher Etizaz Mohsin found a zero-click vulnerability in Orchids, a vibe coding platform claiming around one million users. In a controlled test with a BBC technology journalist in December 2025, Mohsin demonstrated gaining full remote access to the journalist’s laptop without any action from the victim.
The Scale of Exposure Is Growing Fast
AI-generated code now represents 22% of all merged code across the industry, according to late 2025 research. On GitHub, estimates put AI-generated code at 46% of everything written through Copilot. About 75% of developers report using some form of AI-assisted coding. Roughly 63% of vibe-coding users are estimated to be non-developers, including writers, marketers, and students.
Georgia Tech’s Vibe Security Radar, launched in May 2025 to track CVEs tied to AI coding tools, recorded 6 cases in January 2026, 15 in February, and 35 in March. Researchers estimate the true count of AI-generated-code vulnerabilities in the open-source ecosystem sits between 400 and 700 cases, roughly five to ten times the detected figure, because most AI coding tools do not leave identifiable commit metadata.
For founders shipping software to customers, the risk is not theoretical. Security audits and proper configuration reviews are no longer optional steps that can follow launch.
Read more: Read this before you vibe-code another app