Jailbreaking Large Language Models: Techniques, Trends, Defenses, and Open Challenges

Large language models (LLMs) are rapidly becoming embedded in everyday applications, security workflows, and decision-support systems. Along with their capabilities, we have also seen a rise in techniques aimed at bypassing their guardrails. This paper presents a consolidated analysis of recent jailbreaks attacks and groups them into three high-order categories based on their underlying mechanism. Unlike prior surveys, this work introduces a mechanism-grounded taxonomy and an explicit attack-to-defense mapping that unifies cognitive, representational, and system-level jailbreak pathways. We highlight the emerging trends that make modern jailbreaks more effective and discuss where current defenses tend to fail in practice. Building on these insights, we outline practical, engineering-focused best practices to lower jailbreak risk in deployed systems. The paper closes with key open challenges—from benchmark gaps to reproducibility issues and agentic safety—that must be addressed to build more robust and dependable LLM safety architectures.