GPT-4 Can Exploit Most Vulns Just by Reading Threat Advisories

UIUC researchers demonstrated that a GPT-4–based agent, when paired with a ReAct framework and tools, could automatically exploit 13 of 15 tested open-source vulnerabilities (including several high/critical CVEs), highlighting a practical risk that LLMs can accelerate and scale vulnerability exploitation; the report urges organizations to adopt security best practices and consider similar AI capabilities for defense while acknowledging current limitations and no direct evidence of active abuse in the wild.