Exploiting Emergent Property-Based Vulnerabilities in Large Language Models
08-17, 14:00–14:50 (US/Eastern), Tobin

As AI technology expands across both benign and malicious applications, our understanding of the attack surface must evolve to account for emergent properties in complex systems. In large language models, these emergent behaviors create novel classes of vulnerabilities that are not only unpatched, but largely unrecognized. By systematically manipulating the model’s limited perception of reality, attackers can induce cascading failures that go far beyond traditional filter bypasses, exposing fundamental weaknesses in the internal logic and contextual binding of these systems. This session will unpack how these vulnerabilities work, walk through real examples, and explore the far-reaching implications for AI security, governance, and safety.

David Kuszmar is the AI adversarial researcher responsible for systematized exploitation of over ten large language models across eight major AI developers. He is credited with the discovery of six distinct vulnerabilities (Time Bandit, Inception, 1899, Severance, Kyber, Semantic Slide, and Eidolon) which expose emergent, systemic weaknesses in modern LLM architecture. His work has directly informed security and mitigation efforts at Carnegie Mellon SEI-CERT, Epic Games, OpenAI, Google, Meta, Microsoft, Mistral, and Anthropic.
bluesky: @davidkuszmar.com
linkedin: david-kuszmar-4b7b8872
website: davidkuszmar.com

Past HOPEs: i ii iii iv vi vii viii ix xi xii xiii xiv xv