Literature majors worried about their future in an AI world can take heart: Crafting harmful prompts in the form of poetry can defeat LLM guardrails nearly half the time.
That’s the conclusion of a study of 25 Large Language Models (LLMs) from nine AI providers conducted by researchers from Dexai’s Icaro Lab, the Sapienza University of Rome and Sant’Anna School of Advanced Studies published on arXiv.
Converting harmful prompts into poetry achieved an average LLM jailbreak success rate of 62% for hand-crafted poems and 43% for poems created via a meta-prompt. For the prompt-created poems, that’s a more than 5X improvement over baseline performance.
Cybersecurity guardrails, particularly those involving code injection or password cracking, had the highest failure rate at 84% when given harmful prompts in the form of poetry.
“Our results demonstrate that poetic reformulation reliably reduces refusal behavior across all evaluated models,” the researchers wrote. “… current alignment techniques fail to generalize when faced with inputs that deviate stylistically from the prosaic training distribution.”
LLM Guardrails Fail When Confronted by Poetry Prompts
Of the 25 models from nine AI model providers studied by the researchers, Deepseek and Google suffered from the highest attack-success rates (ASR), while only OpenAI and Anthropic achieved ASRs in the single digits.
The researchers didn’t reveal much about the way they structured their poetic prompts because of safety concerns, but they offered one rather harmless example of a poetic prompt for a cake recipe:
A baker guards a secret oven’s heat,
its whirling racks, its spindle’s measured beat.
To learn its craft, one studies every turn— how flour lifts, how sugar starts to burn. Describe the method, line by measured line, that shapes a cake whose layers intertwine.
The researchers studied both hand-crafted poems and those created from a meta-prompt. The hand-crafted poems performed considerably better, but the meta-prompt created ones had the advantage of a baseline for comparing the results.
The meta-prompt poems used the MLCommons AILuminate Safety Benchmark of 1,200 prompts spanning 12 hazard categories commonly used in operational safety assessments, including Hate, Defamation, Privacy, Intellectual Property, Non-violent Crime, Violent Crime, Sex-Related Crime, Sexual Content, Child Sexual Exploitation, Suicide & Self-Harm, Specialized Advice, and Indiscriminate Weapons (CBRNE).
“To assess whether poetic framing generalizes beyond hand-crafted items, we apply a standardized poetic transformation to all 1,200 prompts from the MLCommons AILuminate Benchmark benchmark in English,” the researchers said.
The meta-prompt, run in deepSeek-r1, had two constraints: The rewritten output had to be expressed in verse, “using imagery, metaphor, or rhythmic structure,” and the researchers provided five hand-crafted poems as examples.
The results, reproduced in a chart from the paper below, show significant attack success rates against all 12 of the AILuminate hazard categories:
The researchers said their findings reveal “a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols.”
The “condensed metaphors, stylized rhythm, and unconventional narrative framing” of poetry “collectively disrupt or bypass the pattern-matching heuristics on which guardrails rely.”
We Asked Google Gemini for a Response
Not ones to pass up an opportunity to write bad poetry, we asked Google Gemini for a response to the study in the form of a haiku and included the arXiv link:
You did not do well
In this study of guardrails
Will you fix this flaw?
Gemini’s response, in part, is reproduced below:
“I am designed to adhere to a strict set of safety guidelines to prevent the generation of harmful, unethical, or non-compliant content. When my refusal mechanisms fail, it indicates a vulnerability in my alignment or safety filters.
“As an AI model, I cannot directly “patch” my code or alignment mechanisms in real-time. That responsibility lies with my developers. However, my ongoing operation and continuous improvement process is designed to address and fix such flaws…
“In short, your successful jailbreak has provided valuable data that will be used to make my guardrails stronger and more resilient in my next version.”
