ECE PhD Thesis Defense: Saad Ullah

ECE PhD Thesis Defense: Saad Ullah

Title: From Security Assistants to Self-Improving Security Agents

Presenter: Saad Ullah

Advisor: Professor Gianluca Stringhini

Chair: TBD

Committee: Professor Gianluca Stringhini, Professor Ayșe Coskun, Professor Manuel Egele, Professor Giovanni Vigna

Google Scholar Link: https://scholar.google.com/citations?user=oLVqx8YAAAAJ&hl=en

Abstract: Large language models (LLMs) have rapidly become useful for cybersecurity, but the source of that usefulness has evolved through three stages, and this dissertation advances an LLM at each one. At first, capability came from the frontier model alone: a security task became feasible only once a more powerful model arrived, leaving practitioners to wait for the model that could finally do the job. As the stack around the model matured, the source of capability shifted to the practitioner, who could use their own expertise to design the agentic harness and the task context around an existing standalone model and extract capabilities the standalone model could not deliver on its own. Yet this expertise-driven, manually engineered design does not scale, since every new task demands its own human effort. The natural endpoint moves the source of capability one step further, removing the human from the design loop with an algorithm that lets a model improve itself on a security task. Working through these three levers, i.e., the standalone model, the hand-built agent harness, and the algorithm for self-improvement, this dissertation turns an LLM from an unreliable assistant into a self-improving security agent.