Work

Cat and Mouse in the LLM Era: An Investigation of Jailbreak Attacks and Defenses

Year: 2026
Type: article
Abstract: Large Language Models (LLMs) are susceptible to jailbreak assaults that can circumvent safety alignment and produce destructive outputs, notwithstanding their impressive performance in a variety of natural language processing applications. This paper provides a systematic survey of the attack methods and defense mechanisms of jailbreaking attack for LLMs. Existing attacks fall into three categories: attacks that manipulate input content at the prompt level, attacks that leverage algorithmic search through optimization, and attacks in which artificial intelligence systems act as self-directed adversaries through autonomous reasoning. Then, the defense strategies are reviewed and categorized into prompt-level defenses, model-level defenses, and proactive system-level defenses. The review highlights that while substantial progress has been made, existing defenses are often unable to balance safety with model utility, and generalize poorly to novel attack strategies. This paper also discusses open challenges and promising directions for future research such as layered defense architectures and proactive mechanisms that interfere with the attack process itself.
Author: Puhan Chen
Language: English
Open Access status: hybrid
OpenAlex ID: W7234858657
FWCI:
Cites: 18
Cited by:
Related to: 10