News Multiple ChatGPT instances work together to find and exploit security flaws — teams of LLMs tested by UIUC beat single bots and dedicated software

[Admin]

UIUC computer scientists have innovated a system for teams of LLMs to work together to exploit zero-day security vulnerabilities without an explanation of the flaws, a major improvement in the school's ongoing research.

Multiple ChatGPT instances work together to find and exploit security flaws — teams of LLMs tested by UIUC beat single bots and dedicated software : Read more

 

[Alvar "Miles" Udell]

This is a perfect use for LLMs, they can work faster and in more creative ways than human security researchers can, and are equally as suited for small independent companies as they are Fortune 100.

 

[hotaru251]

technology advancing faster than security is going to be major issue going forward.

 

[Mikiejb]

I don’t buy it. I tasked chat gpt with analyzing a (very) simple spread sheet and it couldn’t even count.

 

[jp7189]

Open source, open weight models (I'm thinking llama3 specifically) can be very good and more flexible in fine tune training compared to closed source models (like chatgpt) which have guardrails on what can be done. I wonder if this system could be even better if the researchers used datasets of specific classes of exploits and used them to fine-tune an open model.

Take for example llama3 vs llama3-lumi-tess. The difference a little fine tuning makes is pretty amazing.

 

[batero]

UIUC computer scientists have innovated a system for teams of LLMs to work together to exploit zero-day security vulnerabilities without an explanation of the flaws, a major improvement in the school's ongoing research.

Multiple ChatGPT instances work together to find and exploit security flaws — teams of LLMs tested by UIUC beat single bots and dedicated software : Read more

This will be the new norm. Model A performs attack, model B respond to attack. Machines will get smarter than humans.

 

[grantorinounforgiven]

The UIUC system leveraging teams of LLMs to identify zero-day vulnerabilities gives us an idea of how AI can revolutionize cybersecurity. While LLMs can struggle with simple tasks, their combined capabilities in this context seem promising. By the way, speaking of Chat GPT for sheets, the extension I'm using performs quite well with tables, charts, lists, and data analysis, no hiccups. I agree with the point about the technology outpacing security measures - it’s a double-edged sword. However, using open-source models like LLaMA3 could further enhance this system’s potential. This innovation could indeed become the new standard in cybersecurity, where AI not only finds but also defends against exploits.