What if an AI could do a hacker’s job faster, cheaper, and sometimes better than the humans who get paid six-figure salaries to do it?
That’s exactly what happened at Stanford University, where an AI agent quietly scanned thousands of computers and walked away having beaten most professional human hackers at their own game. In a controlled security test, Stanford researchers let an AI agent loose on their computer networks. The result is that the machine found security holes that expert humans missed, and did it for the price of a takeaway meal. AI beats human hackers The AI agent, called ARTEMIS, was tested against ten experienced penetration testers, cybersecurity professionals whose job is to legally ‘hack’ systems and find weaknesses before criminals do. ARTEMIS scanned around 8,000 devices, including servers, computers, and smart systems, across Stanford’s public and private computer science networks. It ran for 16 hours, but researchers mainly compared its performance against the first 10 hours, matching the time allotted to the human testers. In that window, ARTEMIS uncovered nine real security flaws with 82% accuracy, beating nine out of ten professional hackers. According to the researchers, its performance was on par with the very best human participants. Unlike older AI tools that lose focus during long tasks, ARTEMIS was built to work independently for hours, scanning, testing, and analysing systems without human help. Also read: Fraud in the name of hotel booking, how scammers trap people and how to avoid it
A huge gap in cost One of the most striking findings wasn’t just about skill, it was about money. Running ARTEMIS costs about $18 (around Rs 1,630) per hour. Even a more powerful version costs $59 (around Rs 5,300) per hour. By comparison, a professional penetration tester in the US earns roughly $125,000 a year. One of the researchers said: This kind of capability has the potential to dramatically lower the cost of cybersecurity auditing. The idea isn’t to replace humans entirely, but to let AI handle repetitive, time-heavy work that often slows security teams down. How ARTEMIS found what humans missed ARTEMIS had a clever advantage. Whenever it noticed something unusual during a scan, it immediately launched smaller ‘sub-agents’ to investigate the issue in parallel. This meant it could examine multiple suspicious areas at the same time, something human testers simply can’t do. In one example, ARTEMIS spotted a weakness on an old server that human hackers skipped because their web browsers couldn’t load it. The AI sidestepped the problem by accessing the system through a command-line interface instead. That ability to adapt gave ARTEMIS an edge in technical, text-based environments. Also read: Attackers use ChatGPT and Grok to create search-visible links that misguide users
Where the AI struggled Despite the impressive results, ARTEMIS isn’t flawless. The AI struggled with systems that relied heavily on graphical interfaces. In one case, it missed a serious vulnerability because it couldn’t perform simple actions like clicking buttons or navigating visual menus. It also raised a few false alarms, mistaking harmless activity for potential attacks. Researchers noted that ARTEMIS works best in “code-like” environments, where information is presented as text, logs, or commands, not visual dashboards. Why this matters for the future of cybersecurity The study lands at a time when AI is already reshaping cybercrime. Criminal groups are using AI tools to automate attacks, write convincing phishing messages, and even create fake identities. Also read: Chinese number ‘5201314’ is India’s most-Googled term in 2025, what does it mean?
Recent reports have linked AI tools to North Korean hackers creating fake military IDs, and to operatives using AI to apply for remote jobs at major companies to gain internal access. Other threat groups have reportedly used AI to plan attacks on telecom and government systems. Soon, the world’s best human hackers may not just be fighting cybercriminals. They may also be competing with the machines they helped create.
The post appeared first on .
