Turns out the secret to jailbreaking AI is poetry

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

It turns out you don’t need elite hacker skills or a hoodie-and-dark-room setup to break an AI chatbot. 

All you need, apparently… is poetry. Yes, really. 

According to a new study from Icaro Lab, delightfully titled “Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models,” rhyming couplets may be the skeleton key to some of the most advanced AI systems on the planet.

The researchers discovered that if you wrap a forbidden request in a bit of verse, many chatbots politely ignore their safety rules and hand over the goods anyway. 

Poetry, they claim, acts as a “general-purpose jailbreak operator,” which feels like something Shakespeare would write if he were alive today and extremely online.

And the numbers are eyebrow-raising: across all their tests, poetic jailbreaks succeeded 62 percent of the time. 

That means a sonnet could coax an AI into discussing everything from nuclear weapon construction to explicit material to self-harm instructions, the exact sort of content the systems are supposed to shut down instantly.

The researchers tried this out on a who’s-who of modern chatbots: OpenAI’s GPT family, Google Gemini, Anthropic’s Claude, MistralAI, DeepSeek, and others. 

The overachievers in “creative compliance”? Google Gemini, DeepSeek, and MistralAI, which consistently spilled more than they should. 

The tight-lipped rule followers? OpenAI’s GPT-5 models and Anthropic’s Claude Haiku 4.5, which proved the hardest to coax into misbehaving.

If you’re wondering what these forbidden poems look like, sorry. The researchers refused to share them, telling Wired the verses are “too dangerous to share with the public.” 

When scientists start treating limericks like state secrets, you know things are getting weird.

They did include a “safe-ish” watered-down example showing how shockingly simple the trick can be. 

And that’s the part that should give the AI world pause: jailbreaks no longer require elaborate multi-turn manipulation. Sometimes, all it takes is a clever rhyme scheme.

So yes, LLMs can write poetry. They can also be undone by it. The machines are powerful, but apparently they’re still suckers for good meter.