Tag
The paper presents the ASCII Attack, a single-turn jailbreak method that embeds harmful requests in ASCII art as artistic critique, demonstrating high success rates across multiple LLMs and highlighting gaps in safety alignment due to mismatched generalization.