AI models have become highly reliable at writing code that compiles, but they are still failing basic security tests in ...