The School of Computer Science at the University of Windsor is pleased to present…
Title: Who Taught the Model? Who Told the Agent? Tracing Malicious Influence from Fine-Tuning to Tool Use
Colloquium Presentation by: Dr. Birhanu Eshete
Date: Friday, 6 November 2026 Time: 10:00 – 11:00 am
Location: Erie Hall, room 3123
Abstract: Modern AI systems increasingly depend on information they do not fully control. During fine-tuning, LLMs learn from datasets assembled from external sources. During deployment, AI agents retrieve web pages, emails, documents, database records, and other content before making decisions and invoking tools. These capabilities create two seemingly different security problems: malicious training examples can quietly teach a model an attacker-chosen behavior, while malicious external content can redirect an otherwise benign AI agent toward an unauthorized action. In this talk, I will argue that these problems share a deeper structure: both are failures of malicious influence propagation. The key security question is therefore not simply whether an input looks suspicious, but whether we can determine where consequential behavior came from, how untrusted influence propagated through the AI system, and where that influence can be interrupted. I will develop this perspective through two recent lines of work. First, PoisonGraph addresses backdoor poisoning before LLM fine-tuning by examining the optimization influence of individual training examples and using graph structure to identify groups that repeatedly push the model toward shared behavior. These systems point toward a broader direction for AI security: treating provenance and influence tracing as security primitives for understanding, detecting, and interrupting malicious behavior as AI systems learn, reason, and act.
Keywords: AI Security, Large Language Models, Agentic AI
Biography: Birhanu Eshete is an Associate Professor of Computer Science at the University of Michigan-Dearborn, where he directs the Data-Driven Security and Privacy Lab and the Center for Cybersecurity Education, Research, and Outreach. His research develops methods and systems to identify, characterize, and mitigate security, privacy, safety, transparency, and ethical risks in AI systems with emphasis on high-stakes applications such as autonomous vehicles, predictive diagnostics, financial forecasting, and cyber-attack detection. Dr. Eshete’s research has been published in all leading venues in security, privacy, and AI, including IEEE S&P, ACM CCS, USENIX Security, ISOC NDSS, and IEEE/IFIP DSN, and featured in widely accessible venues such as Science Magazine. His research has also contributed to national efforts, including the U.S. National Institute of Standards and Technology (NIST) Trustworthy & Responsible AI Resource Center. He is the recipient of the 2024–2025 Fulbright U.S. Scholar Award, the 2024–2025 Faculty Excellence in Research Award from the College of Engineering and Computer Science, the 2023 NSF CAREER Award from the U.S. National Foundation, the 2018 USENIX Security Symposium Distinguished Paper Award, and was a finalist for the Best Applied Security Research Award in North America in 2018.
