Cybersecurity with and against LLM Agents: Capabilities, Evaluation, and Countermeasures

PhD Qualifying Examination


Title: "Cybersecurity with and against LLM Agents: Capabilities, Evaluation, 
and Countermeasures"

by

Mr. Kaikai ZHANG


Abstract:

Large language model (LLM) agents are increasingly capable of performing 
complex cybersecurity tasks, creating opportunities for security automation 
while raising new defensive challenges. This survey examines three connected 
questions: how to evaluate the capability boundaries of LLM cyber agents as 
existing benchmarks become saturated, how these capabilities can be improved, 
and how systems can be protected against increasingly capable AI attackers.

We review evaluation methodologies, capability improvement through elicitation 
and training, and AI-native countermeasures. The analysis identifies a shared 
bottleneck: realistic, executable, verifiable, and contained environments that 
support both meaningful evaluation and effective training. It also highlights 
the need to evaluate defenses against adaptive attackers and to measure 
protection through security outcomes. Building on these findings, the proposed 
doctoral research will develop more realistic benchmarks, investigate 
environment-based agent training, and evaluate AI-native defenses under 
attacker adaptation. It will further explore attacker-defender self-play as a 
framework for jointly advancing agent capabilities and system protection.


Date:                   Monday, 12 October 2026

Time:                   4:00pm - 6:00pm

Venue:                  Room 5504
                        Lifts 25/26

Committee Members:      Dr. Dongdong She (Supervisor)
                        Dr. Shuai Wang (Chairperson)
                        Dr. Chaojian Li