AI agent finds more security flaws than human hackers at Stanford (scienceclock.com)

In a groundbreaking experiment at Stanford University, the AI agent ARTEMIS identified more security vulnerabilities than a team of seasoned human hackers within a controlled environment. Operating autonomously for 16 hours across the university's network of approximately 8,000 devices, ARTEMIS found nine valid security flaws with an impressive 82 percent accuracy rate. It outperformed nine out of ten human penetration testers, showcasing its capacity to efficiently tackle complex security tasks that many AI tools struggle with. ARTEMIS stands out due to its innovative approach, utilizing multiple "sub-agents" to investigate vulnerabilities simultaneously, a capability that allows it to cover more ground than human testers, who must handle tasks sequentially. Remarkably, operating ARTEMIS costs merely $18 per hour, considerably less than the average annual salary for human penetration testers in the U.S. While the AI exhibits limitations, such as difficulties with graphical user interfaces and a higher rate of false positives, its results signal a shift in cybersecurity strategies, suggesting that AI could soon play a critical role in vulnerability assessments, rivaling human experts. As the landscape of cyber threats evolves, ARTEMIS demonstrates that AI is no longer just an auxiliary tool but a formidable contender in cybersecurity.