The 2026 AI Code Security Report

Between January and April 2026, Sherlock Forensics conducted security assessments on dozens of applications built with AI coding tools including GitHub Copilot, Claude, ChatGPT and Cursor. The findings were consistent and alarming: the vast majority of AI-generated codebases contain vulnerabilities that would be considered unacceptable in any production environment.

AI code assistants optimize for functionality, speed and developer satisfaction. Security is a constraint that conflicts with those goals. The result is code that works, compiles, passes basic tests and ships to production carrying exploitable vulnerabilities that traditional code review rarely catches because the reviewer did not write the code.

This report presents aggregate, anonymized findings from those assessments. Every statistic reflects real vulnerabilities found in real applications serving real users. The purpose is not to discourage AI-assisted development but to quantify the security gap so that teams can address it before attackers do.

Based on anonymized, aggregate findings from Sherlock Forensics security assessments conducted between January and April 2026. All data has been stripped of identifying information. No individual client or application can be identified from the statistics presented.

Assessments covered web applications, APIs, SaaS platforms and internal tools built using AI coding assistants. Each assessment followed Sherlock Forensics' standard methodology mapped to OWASP Top 10 and MITRE ATT&CK frameworks. Manual testing was performed on every engagement alongside automated scanning.

Applications ranged from pre-launch MVPs to production systems with thousands of active users. The majority were built using Cursor, GitHub Copilot, ChatGPT or Claude as the primary code generation tool.