LLMs Hit Security Plateau: Why AI Code Can't Be Trusted Yet
TechTarget, Tuesday, August 19th, 2025
Frontier models have exhausted quality public code, so synthetic training data recycles the same flaws.
AI-generated code has plateaued on security because frontier models have consumed most high-quality public code and now train on synthetic output that recycles existing weaknesses.
The recurring defects are familiar ones: missing validation, unsafe queries, and incomplete access controls.
The structural problem is that the most secure enterprise code is private and never enters training data, so models never learn from the practices that actually hold up in production.
Meanwhile AI generates vulnerabilities at a higher rate than human developers while security remains unoptimized in model design. The recommendation is to treat AI-generated code as untrusted input requiring independent verification, the same way a third-party component would be treated.