By using large language models to compare software code before and after an update, attackers are shrinking the window for organisations to patch known flaws