Jailbreaking Aligned Large Language Models: A Systematic Survey of Attack Taxonomies, Evaluation Frameworks, and Defense Mechanisms
Year: 2026
Type: preprint
Abstract: Large language models (LLMs) are now deployed to hundreds of millions of users, and their safe behavior depends on alignment techniques such as supervised fine-tuning and reinforcement learning from human feedback. Yet a class of adversarial inputs known as jailbreaks continues to circumvent these safeguards, eliciting outputs the models were trained to refuse. This paper presents a systematic survey of the jailbreak literature. We formalize the threat model, propose a two-dimensional taxonomy that separates attacks by adversary capability (white-box versus black-box) and by exploitation mechanism, and review representative methods in each category, from gradient-based suffix optimization to persuasion-driven and many-shot techniques. We then examine how the community evaluates attacks and defenses, covering benchmarks, success metrics, and automated judges, and we organize defense strategies into prompt-level, model-level, and system-level interventions. Finally, we identify open challenges — including the generalization of defenses, the security of agentic and multimodal systems, and the limits of current evaluation — and outline directions for future research.
Author: Mahmoudpour, Aref
Institution: Islamic Azad University, Tehran
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222872205
Cites:
Cited by:
Related to: