PUZZLED: Jailbreaking LLMs through Word-Based Puzzles

ACL ARR 2026 January Submission7796 Authors

Keywords: Large Language Models, Jailbreak, AI Safety, Alignment, Prompt Engineering

Abstract: As LLM deployment expands, jailbreak attacks pose growing safety concerns. We propose PUZZLED, a novel jailbreak method that exploits LLMs’ reasoning by masking harmful keywords as word puzzles. Using word search, anagram, and crossword puzzles, the model must solve the puzzle to reconstruct the harmful instruction before responding. Evaluations on five state-of-the-art LLMs show a high average attack success rate of 88.8\%, including 96.5\% on GPT-4.1 and 92.3\% on Claude 3.7 Sonnet, demonstrating that PUZZLED is a simple yet effective reasoning-based jailbreak strategy.

Paper Type: Long

Research Area: Safety and Alignment in LLMs

Research Area Keywords: adversarial attacks/examples/training, robustness, safety and alignment, red teaming

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7796