AI Models Write Code with Security Flaws 18–50% of the Time, New Study Finds | Not Hacker News!