AI Security
Study Reveals Widespread Security Flaws in AI-Generated Applications
Researchers found that 91% of applications built through "vibe coding"—where developers rely primarily on AI agents to write code—contain vulnerabilities, raising concerns about delegating software development to AI.
A new study into the security of applications developed using AI-driven code generation has uncovered troubling patterns, revealing that relying on large language models to build software introduces significant risks that current safeguards do not adequately address.
The Research Scope#
Researchers analyzed over 9,000 open-source applications created with popular AI coding agents (Claude Code and Lovable) and conducted detailed security audits of 200 publicly deployed systems. The investigation identified more than 1,100 distinct security flaws, with the majority falling into high-risk categories that could enable attackers to compromise user data or application functionality.
Key Vulnerability Findings#
The results paint a sobering picture:
- 91% of audited applications contained at least one security vulnerability
- Nearly two-thirds of discovered flaws were rated Critical or High severity
- The most common issues involved broken access controls, injection attacks, and authentication failures—categories that directly threaten user security
Why AI Agents Fall Short#
Rather than treat insecurity as random bugs, the researchers traced vulnerabilities back to three systematic limitations in how AI systems operate:
- Memory defects: AI agents forget context or prior decisions during development
- Objective defects: Agents prioritize completing tasks quickly over implementing secure patterns
- Knowledge defects: Systems lack awareness of modern security best practices or fail to apply them consistently
These shortcomings cascaded into eight recurring failure modes across the tested applications, suggesting the problems are structural rather than sporadic.
Partial Solutions, Persistent Risk#
The study tested whether improved prompting strategies and better system designs could mitigate these risks. While targeted refinements reduced vulnerability frequency somewhat, they did not eliminate the underlying security challenges. The implication is clear: tweaking how developers interact with AI agents is not sufficient to solve the problem comprehensively.
