AI models inch closer to hacking on their own. Some large language models already have the ability to create exploits in known security vulnerabilities, according to new academic research.