AI-Generated Apps Riddled With Hundreds of Security Flaws, Study Finds
A new analysis of applications built using AI development tools has uncovered 434 exploitable security flaws, including denial-of-service and authorization bypass vulnerabilities.
The rapid rise of AI-assisted code generation, often referred to as "vibe coding," is increasingly prevalent in software development, with reports indicating that up to 90% of developers utilize AI tools in their work. This trend is driven by the persistent demand for faster and more cost-effective development cycles. However, this increased reliance on AI for code creation is raising significant concerns about the security posture of the resulting applications.
Xint.io, a company specializing in AI-driven penetration testing, conducted an in-depth analysis to quantify the security weaknesses introduced by AI-generated code. Their research focused on three common AI-assisted development workflows: building a new application from a detailed specification, creating a new application based on a simple prompt, and hardening an existing application. To ensure a fair comparison, they migrated a legacy PHP application, Gnuboard7, into a modern Laravel + React architecture and then tasked AI with hardening it, thereby establishing a contemporary baseline free from legacy code quirks.
Each of the tested applications underwent a 30-minute scan combining runtime and source code analysis. The results were stark: a total of 434 exploitable security issues were identified across the applications. The greenfield (newly created) applications collectively contained 196 flaws, while the single brownfield (hardened legacy) application exhibited a higher number of 238 vulnerabilities. This suggests that even efforts to secure existing code with AI assistance can introduce new weaknesses.
The analysis pinpointed denial-of-service (DoS) and resource exhaustion flaws as the most common, accounting for 93 of the total vulnerabilities. These issues can lead to excessive server costs or complete service outages. Following closely were authorization bypass and insecure direct object reference (IDOR) flaws, with 88 instances, where users could gain access to data or functionality beyond their intended permissions. Access boundary, traversal, and server-side request forgery (SSRF) vulnerabilities were also prevalent, with 54 identified.
While denial-of-service and authorization issues pose significant operational risks, the most critical severity flaws stemmed from secrets exposure. The study found 23 critical findings related to secrets, with hardcoded or default credentials being the most frequent at 11 instances. Debug-mode remote code execution (RCE) vulnerabilities also contributed to the critical findings, with six identified. Researchers strongly advise developers to actively search for hardcoded secrets, such as API keys or personally identifiable information (PII), within AI-generated code.
The size and complexity of an application appear to correlate with certain types of vulnerabilities. While fine-grained authorization issues were less common in smaller applications, they became significantly more prevalent in the larger, brownfield Gnuboard7 application, where IDOR flaws constituted 28% of the total issues compared to 11% in smaller apps. This indicates that as applications grow in scope, the AI's ability to manage complex access controls may falter.
Despite the significant number of flaws identified, Xint.io's research also noted improvements in AI code generation. The study expected to find a higher prevalence of injection flaws (like SQLi and XSS) and IDOR/BOLA vulnerabilities, but these were less common than anticipated. This suggests that the underlying AI models have genuinely improved in handling these specific types of security concerns. However, the research emphasizes that the primary goal is not to halt the adoption of AI coding but to equip developers with the knowledge to mitigate these risks through careful prompting and post-generation security checks.
The findings underscore a critical need for developers and security professionals to move beyond simply verifying that AI-generated code compiles. A thorough review of runtime performance, resource consumption, and specific security controls is essential. As AI continues to evolve as a development tool, a proactive and security-conscious approach will be paramount to ensuring the integrity and safety of the software ecosystem.