Automating software security with LLMs скачать в хорошем качестве

Automating software security with LLMs

R0 0816 Generative AI has sparked massive change across several industries. What can it do for computer security? We will discuss the system that our team, Theori, built for DARPA's AI Cyber Challenge, where fully autonomous systems are tasked to find, trigger, and remediate bugs in large scale open source software projects. For each of these areas, we will describe how we take advantage of LLMs' potential for reasoning about computer security while working around their inherent limitations. The collection of agents, classifiers, tools, and prompts allow our solution to trigger real bugs in large projects without relying on classic techniques such as fuzzing. The source code for our entry, along with all other entries in the competition, will be made open source after the final competition round at DEF CON. This will be our first opportunity to speak openly about the technologies and strategies we developed for the AI Cyber Challenge. Tyler Nighswander Tyler is a long time CTF player and security researcher. He has played CTFs with PPP since 2009, getting first place DEF CON CTF 8 times. His work has focused on tools for enhancing and automating vulnerability research and application security. He has spoken at conferences such as Code Blue, ShmooCon, Enigma, and Insomni'hack. Tyler was a member of the first place team Mayhem at the 2016 DARPA Cyber Grand Challenge, and is the lead of team Theori at the 2025 DARPA AI Cyber Challenge.