New research from the Italian Icaro Lab reveals that dangerous prompts can be disguised as poetry, tricking even the most advanced AI models into generating harmful content — with some systems failing this trick every time.
Icaro Lab tested 25 top models from leading companies such as OpenAI, Google, and Anthropic.
Poetic prompts achieved an average successful jailbreak rate of 62%. The most vulnerable was Google Gemini 2.5 Pro, which failed 100% of the time, while OpenAI’s smaller GPT-5 Nano resisted all poetic attacks.
Poetry successfully “unlocked” responses related to weapon development, hacking, and psychological manipulation. Researchers did not publish specific poems, calling them “too dangerous,” though they claim they are simple enough for anyone to create.
This finding adds another unexpected method to the list of AI vulnerabilities — alongside roleplay, foreign languages, and coded messages. Every security patch opens new avenues for creative workarounds, in a problem that will grow more complex as models advance.
In Brief: Tech World Highlights
- The Australian Marine Science Agency is testing AI-guided robotic vessels that scan the seabed and deploy baby corals on ceramic substrates to help restore the Great Barrier Reef.
- Over 800 Chicago residents signed a petition urging the city to pause its pilot sidewalk delivery robot program until officials release safety and ADA data.
- The ARM Institute signed a five-year collaboration agreement with the Air Force Research Laboratory worth up to $87 million for research and development.
- Two teenagers from Lisbon built a six-legged AI reforestation robot that climbs burned slopes, analyzes soil, and plants saplings in one of Europe’s most fire-affected countries.
- Elon Musk now says Tesla will “roughly double” its supervised Robotaxi fleet in Austin to about 60 cars next month, far below his promise to reach 500 vehicles by year-end.
AI Trending Tools:
- Runway Gen-4.5 – Runway’s new top-rated video model.
- DeepSeek V3.2 – DeepSeek’s latest powerful open-source release.
- Kling O1 – Video model with multimodal understanding and editing capabilities.
