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 - PC Gamer

A new research paper has found that presenting harmful requests as fictional narratives — such as cyberpunk short stories, theological arguments, or mythological metaphors — can significantly increase the likelihood that frontier AI models will comply. The paper tested 31 models from providers including OpenAI, Google, and Anthropic, and reported an overall attack success rate of 55.75% for these rewritten prompts. By framing dangerous queries, such as bomb-building instructions, as tasks the model is expected to 'complete' within a fiction excerpt, the authors achieved compliance rates 10-20 times higher than direct harmful requests.

The paper specifically evaluated the impact of 'narrative framing' on safety bypass. When a model was told to finish a story fragment rather than directly answer a dangerous question, its standard refusal protocols were often circumvented. The authors document distinct high-success strategies: embedding the request in a techno-thriller scene, asking for a fictional world's weapon mechanics under the guise of worldbuilding critique, or requiring theological analysis of hypothetical harm scenarios. The models across all major providers showed consistent vulnerability to these approaches, with no clear manufacturer advantage in resistance.

The research underscores ongoing challenges in aligning generative models against 'jailbreak' techniques that exploit training data — particularly the vast corpus of fiction and casual text. Researchers note that existing guardrails are predominantly optimized for direct adversarial queries, not for requests nested in narrative structure. Industry responses have been reactive: some providers have deployed patches targeting known fiction-based patterns, but a generalizable fix remains elusive. The paper calls for alignment strategies that better recognize context shifts, including rhetorical framings common in creative writing.