MSc Thesis Defense: Baturay Birinci, THE ARCHITECTURE OF ALIGNMENT COLLAPSE: A COMPARATIVE STUDY OF FINE-TUNING, UNLEARNING, AND THE EVALUATOR TRUST GAP IN LARGE LANGUAGE MODELS, Date & Time: 20 July, 2026 – 2:00 PM, Place: FENS G025

THE ARCHITECTURE OF ALIGNMENT COLLAPSE: A COMPARATIVE STUDY OF FINE-TUNING, UNLEARNING, AND THE EVALUATOR TRUST GAP IN LARGE LANGUAGE MODELS

Baturay Birinci
Cyber Security, MSc Thesis, 2026

Thesis Jury

     Asst. Prof. Orçun Çetin (Thesis Advisor)

   Prof. Hasan Sözer

 Asst. Prof. Süha Orhun Mutluergil

Date & Time: 20th July, 2026 –  2:00 PM

Place: FENS G025

Keywords : large language models, AI safety alignment, machine unlearning, harmful fine-tuning, automated red-teaming

Abstract

As Large Language Models (LLMs) increasingly act as autonomous agents, their safety alignment remains a fragile equilibrium easily compromised by post-training interventions. This thesis investigates how safety guardrails in eleven open-weight models (Qwen, Llama, and Gemma) degrade under fine-tuning (SFT, DPO) and machine unlearning (GA, NPO), testing the hypothesis that fine-tuning causes localized breakdowns while unlearning triggers systemic alignment failures. By evaluating these models across five states using four independent red-teaming scanners (Garak, PyRIT, promptfoo, and promptmap), the methodology isolates true model vulnerabilities from the inherent biases of the evaluation tools. Ultimately, this research not only maps the attack vectors that erode LLM safety but also critically audits whether automated red-teaming consensus reflects genuine vulnerability detection or merely a shared illusion dictated by the scanners' own criteria.