Cybercriminals have begun exploiting AI-powered email security systems using a technique called indirect prompt injection, according to new research from Sublime Security.