SecAppDev 2024 lecture details
Vulnerabilities of Large Language Model Applications
The session will start with a quick primer on data-driven AI and the key mechanisms behind LLMs. Then we will explore the general threat landscape, including academic attacks and more practical threats (OWASP Top 10 for LLMs).
Wednesday June 5th, 11:00 - 12:30
Room West Wing
Download handoutsAbstract
Large Language Models (LLMs) have recently emerged as a transformative technology with a potential to affect every industry. While the internal workings of LLMs are not entirely understood even by their creators, their rapid adoption has already revealed alarming failures.
In this lecture, we will overview the complex interplay of previously known and newly introduced vulnerabilities underpinning real-world LLM applications. The goal is to raise awareness and move towards a fundamental understanding of what it might take to ensure privacy and security of this fast-evolving ecosystem.
Key takeaway
LLMs are a vulnerable intermediary between users and information. Increasing autonomy, complexity and integration of AI amplifies all existing risks.
Content level
Deep-dive
Target audience
Developers, industry professionals, technology executives, policy makers, educators
Prerequisites
Participants with varying levels of expertise can gain valuable insights. Session "AI Security: Essentials to Advanced" is a recommended prerequisite.
Vera Rimmer
Research expert, DistriNet, KU Leuven
Expertise: Computer security and privacy, applied machine learning and deep learning
Related lectures
AI Security: Essentials to Advanced
Introductory lecture by Jim Manico in room Lemaire
Monday June 3rd, 16:00 - 17:30
Unpack AI security: business impacts, ethics, LLM challenges, privacy, and regulations like the EU AI Act. Essential for secure AI deployment.
Key takeaway: Secure and ethical AI deployment requires understanding risks, regulations, and best practices in technology and governance.
When network protocols meet new threat models
Introductory lecture by Mathy Vanhoef in room West Wing
Wednesday June 5th, 09:00 - 10:30
This presentation will argue that several past (wireless) protocol attacks were found by creatively thinking about threat models.
Key takeaway: Attacks only get better: either by finding new flaws or by introducing new threat models.
Winning the war in cyber
Keynote lecture by Jessica Robinson in room Lemaire
Monday June 3rd, 09:15 - 10:30
How well we adapt continues to influence our security strategies, our creativity, and our culture, in our companies and in our industry. It seems starting with ourselves is a natural place to begin.
Key takeaway: What the evolution of the security practitioner, and leader, will look in the future in winning the daily battles in cybersecurity.