GPT-4 can exploit vulnerabilities by reading CVEs

OpenAI's GPT-4 has been shown to autonomously exploit real security vulnerabilities by interpreting CVE advisories, outperforming other models and vulnerabilty scanners.

• OpenAI's GPT-4 can exploit 87% of vulnerabilities
• UIUC researchers tested LLM on real systems
• LLM agents simplify the exploitation process
• Future models may be more capable
• Study raises implications for security practices