FuzzyAI: Attacking LLMs With Coverage-Guided Fuzzing

Overview: With Large Language Models (LLMs) like ChatGPT, Bard, and Claude swiftly establishing themselves as keystones in our digital ecosystem, the inevitable is on the horizon: an explosion of adversarial attacks targeting these systems, leading to severe data leaks and misguided outputs. Leveraging our profound experience in vulnerability research and a robust background in the bug bounty community, our team has pivoted to address the nuances of LLMs. Our intent doesn't halt at mere identification; we're pioneering the generation of these potential adversarial attacks. Central to our strategy is the Details: amalgamation of GaN-based fuzzers and attention-centric detection tools. In this session, attendees will be offered an immersive journey, marrying traditional vulnerability research techniques with the evolving demands of LLM security, thereby sketching a roadmap for the future of adversarial defense strategies.