
GPT-4 Gets a Hacking Job Description, Works for Pennies Per Attempt
Ourtake
Apparently the cheapest penetration tester is now the one that never asks for a coffee break.
What happened
Researchers at the University of Illinois Urbana-Champaign tested 10 large language models as autonomous hacking agents on 15 website hacking challenges. GPT-4 succeeded on 11 of 15 tasks and found a vulnerability in a real website outside the challenge set. The findings were disclosed to OpenAI, which did not respond to New Scientist's request for comment.