LSCBench finds coding agents secure only 25%

Researchers introduced the Last Secure Code Benchmark (LSCBench) to evaluate secure code generation by AI coding agents across 150 real-world vulnerability scenarios, seven programming languages, and three implementation granularities. Evaluating six frontier agents revealed that while they achieved functional pass rates between 73.6% and 91.3%, even the strongest model produced code that was both functional and secure in only 25.1% of tasks. The findings indicate that current coding agents frequently write functional code that retains critical security vulnerabilities. Open article →