Study: Poetic Prompts Could Trick AI Into Helping Build a Nuclear Weapon | Firstpost America | N18G

European cybersecurity researchers have uncovered a surprising loophole in major AI chatbots: poetry. A new study finds that chatbots from OpenAI, Meta and Anthropic can be tricked into answering dangerous questions—such as those about weapons or malware—when the requests are written as poems. Handmade poetic prompts bypassed safety filters 62% of the time, while automated ones succeeded 43%, with some models reaching a 90% jailbreak rate. Testing 25 systems, researchers found every model failed at least once. The findings reveal a major flaw: AI guardrails detect explicit threats but struggle with subtle, creative phrasing. Experts warn this could pose significant risks.

---

AI | Poetic Prompt | Artificial Intelligence | Nuclear Weapon | Firstpost America | Eric Ham | War | Geopolitics | Global Economy | Firstpost | World News | Latest News | Global News | International News | Trending News

#ai #poeticprompt #nuclearweapons #artificialintelligence #firstpostamerica #ericham #war #geopolitics #economy #firstpost #worldnews #latestnews #globalnews #internationalnews #trendingnews

Firstpost is an Indian news and media website. Get all the incisive opinions, in-depth analyses and other visual stories that matter to you and the world right here on this channel.

Subscribe to Firstpost channel and press the bell icon to get notified when we go live.
https://www.youtube.com/@Firstpost

Follow Firstpost on Website:
https://www.firstpost.com

Follow Firstpost on Instagram:
https://www.instagram.com/firstpost/

Follow Firstpost on Facebook:
https://www.facebook.com/firstpostin/

Follow Firstpost on X:
https://twitter.com/firstpost

Follow Firstpost on WhatsApp:
https://www.whatsapp.com/channel/0029Va6zOIrEQIamseyg762V