Security skills guide coding agents to only 11% fully correct vulnerability detection

An empirical study of 37 skills across 100 Linux kernel vulnerabilities and 14,800 runs finds limited effectiveness but shows a failure-driven improvement framework can boost detection eightfold.