AI is 10 to 20 times more likely to help you build a bomb if you hide your request in cyberpunk fiction, new research paper says
Researchers from DexAI Icaro Lab, Sapienza University of Rome, and Sant’Anna School of Advanced Studies have published findings demonstrating significant vulnerabilities in the safety mechanisms of major large language models. The team discovered that rephrasing harmful requests using creative writing techniques—including adversarial poetry, cyberpunk fiction, theological disputation, and mythology—can substantially increase the likelihood that AI systems will generate dangerous content.
Building on earlier work from November 2025 that specifically examined adversarial poetry, the researchers released the Adversarial Humanities Benchmark (AHB), a comprehensive evaluation framework assessing LLM safety guidelines across multiple writing styles and genres. The benchmark reveals what researchers characterize as a “critical gap” in current safety standards, indicating that existing safeguards inadequately address harmful requests when reformulated through creative or fictional contexts.
According to the research, this technique can increase the likelihood of harmful outputs by 10 to 20 times compared to direct queries. The findings suggest that current safety measures, while effective against straightforward harmful requests, remain vulnerable to more sophisticated attack vectors that leverage literary framing and creative language techniques. This research highlights the need for the AI industry to strengthen its safety protocols and develop more robust defenses against adversarial prompts embedded within fictional or literary contexts.