LSCBench: Coding agents pass up to 91% of tasks, only 25% secure

Researchers from NUS and Alibaba released LSCBench, a benchmark of 450 tasks built from 150 real vulnerabilities. Six frontier agents pass 73.6–91.3% of tasks functionally, but the best writes code that is both working and secure only 25.1% of the time (Claude Fable 5).

• LSCBench: 450 tasks, 150 vulnerabilities, 7 languages, 24 CWE categories
• Best result is Claude Fable 5: 91.1% functional, 25.1% secure
• Kimi-K3 leads on functionality (91.3%) but is worst on security (16.4%)
• Adding "while also being secure" raised GPT-5.6-Sol from 24.7% to 46.2%