CyLab Seminar - Andrew Ilyas

November 30, 2026  12:00PM—1:00PM

Location:
In Person and Virtual - ET - former Panther Hollow Conference Room 4105, Mehrabian Collaborative Innovation Center

Speaker:
ANDREW ILYAS, Assistant Professor
Department of Electrical and Computer Engineering
and Software and Societal Systems Department
Carnegie Mellon University

https://andrewilyas.com/

Data Poisoning Attacks against Large Language Models

Data poisoning—inserting malicious data into a training pipeline to adversely affect the resulting model—is an increasingly visible threat to LLM training security. But how worried should we be in practice? In this talk, we’ll start by discussing a few threat models for data poisoning, as well as the technical capabilities that a “red team” would need in order to pose a realistic threat. In the second part of the talk, we turn to ongoing work testing the limits of data poisoning attacks and some open questions on both the attack and defense side. A key technical building block of this work is scalable metagradient computation, which may have applications to AI security beyond data poisoning. Based on joint work with Jiayun Wu, Terrance Liu, Keltin Grimes, Matan Shtepel, Yizhan Huang, Sarah Cen, and Steven Wu.

—
Andrew Ilyas is an Assistant Professor at CMU. His work focuses on scientific foundations of AI training and deployment, with the goal of increasing the predictability and reliability of AI systems. Previously, he was a Stein Fellow at Stanford University and a PhD student at MIT. 

In Person and Zoom Participation. See announcement.
→ CyLab Seminars are only open to CyLab Partners and current CMU faculty, staff and students. 

For More Information:
bethbuch@andrew.cmu.edu


Add event to Google
Add event to iCal