OPENAI'S GPT-4 CAN EXPLOIT REAL VULNERABILITIES BY READING SECURITY ADVISORIES

Summary from the TLDR InfoSec newsletter

In a new paper, researchers from the University of Illinois Urbana-Champaign claim that AI agents combining large language models like OpenAI's GPT-4 with automation software can successfully exploit real-world security vulnerabilities by reading security advisories. They tested GPT-4 on 15 one-day vulnerabilities and it could exploit 87% of them, outperforming other models and open-source scanners.