LLM Unlearning for Cyber Defense — survey of methods and threats
LLMs deployed in healthcare, finance, and decision support retain sensitive data, copyrighted material, and hazardous knowledge across billions of parameters long after training.
This exposure leaves models vulnerable to extraction, jailbreak, membership inference, and regulatory breach — no longer theoretical risk.
arXiv survey maps unlearning methods, implementation challenges, and real-world attack vectors against deployed systems.