![DALL·E-2024-04-24-16.28.32---A-newspaper-cartoon,-drawn-with-a-pen,-depicting-a-robot-computer-hacker.-The-robot-is-whimsically-styled-with-exaggerated-features-such-as-an-oversiz-u6pbi](https://unicorn-s3.b-cdn.net/DALL·E-2024-04-24-16.28.32---A-newspaper-cartoon,-drawn-with-a-pen,-depicting-a-robot-computer-hacker.-The-robot-is-whimsically-styled-with-exaggerated-features-such-as-an-oversiz-u6pbi.jpg)
How much does it cost in compute for GPT4 to exploit a one-day vulnerability? A mere $8.80.
Now, consider the salary it takes to hire a human cybersecurity expert to deal with that same vulnerability.
Researchers at the University of Illinois Urbana-Champaign have developed AI agents powered by GPT-4 that can autonomously exploit cybersecurity vulnerabilities with alarming effectiveness.
However, there's a catch: these AI agents heavily rely on being provided with CVE descriptions to be successful. Without this information, their success rate drops drastically from 87% to just 7%.
So no Skynet yet, but in the long run, could this be a classic case of creating a monster to fight our battles, only to have it turn against us?
The full study - https://arxiv.org/pdf/2404.08144.pdf