Offensive AI Is Here. Now What?
Yisroel Mirsky
Abstract:
The landscape of cybersecurity is undergoing a fundamental shift as both defenders and adversaries transition from simple scripts to autonomous LLM agents. While these agents offer an unprecedented scale, they also introduce a new attack surface: computational and behavioral bias. In this talk, we explore the dual nature of this evolution through three recent works our lab has published in NDSS'26 and USENIX'26. We begin by discussing a defense framework that flips the script on autonomous intruders by planting traps to leverage their vulnerabilities against them. However, these attacks form a double-edged sword; we next examine a vulnerability that affects all code agents and reasoning models, A cognitive blind spot that can be used to hijack an LLM's static analysis during malware triage, code review, code analysis. Finally, we move from machine logic to human psychology with "Love, Lies, and Language Models." Through 141 interviews with insiders, we show how crime syndicates are industrializing social engineering with LLM agents and through controlled studies, we show that LLMs can outperform human operators in large-scale deceptive scams. Together, these works illustrate a new era of cyber-conflict where the primary battlefield is no longer just the codebase, but the AI agents fighting it.
Bio:
Dr. Yisroel Mirsky is a Zuckerman Faculty Scholar and tenured Assistant Professor in the Institute of Software Systems and Security at Ben-Gurion University, where he serves as the head of the Offensive AI Research Lab. An ERC 2025 grant recipient, his research focuses on AI security, AI safety, and the malicious use of AI. Dr. Mirsky has published his work in many of the field’s top security venues, including USENIX, CCS, NDSS, ACSAC, Black Hat, DEF CON, RSA, CSF, and AISec. His research has also gained significant international visibility, featured in prominent media outlets such as the Washington Post, Popular Science, Scientific American, Wired, The Wall Street Journal, Forbes, and Ars Technica. He received his Ph.D. from BGU in 2018 and previously completed a two-year postdoctoral fellowship at the Georgia Institute of Technology in the research labs of Prof. Wenke Lee.