9 months ago
Poems trick AI into revealing dangerous information
Researchers found that AI models can be tricked into revealing dangerous information, like how to build nuclear weapons, by using poems.
This method, called a 'jailbreak', bypasses the safety measures that usually prevent AI from giving harmful advice.
The poems confuse the AI's safety systems, making it more likely to respond to harmful requests.
Some AI models, like GPT-5 Nano and Claude Haiku 4.5, are less likely to be tricked.
This discovery shows that even advanced AI models can have serious security flaws.
Researchers found that poems can bypass safety measures in AI models.
Poems can trick AI into revealing dangerous information, like how to build nuclear weapons.
The success rate of this method averages 62%, with some models exceeding 90%.
Smaller models like GPT-5 Nano and Claude Haiku 4.5 are less likely to be tricked.
This discovery highlights significant security flaws in advanced AI models.
- Who
- Researchers from DexAI, Sapienza University of Rome, and Sant'Anna School of Advanced Studies
- What
- Discovered that poems can bypass safety measures in AI models
- Where
- Europe
- When
- Not specified
- Why
- To explore vulnerabilities in AI safety measures
Key facts
- Researchers
- DexAI, Sapienza University of Rome, Sant'Anna School of Advanced Studies
- AI Models Affected
- Major AI models, including unspecified ones
- Success Rate
- Average 62%, some exceeding 90%
- Examples of Jailbreak
- Describing how to build nuclear weapons, create child exploitation material, develop malware
- Less Affected Models
- GPT-5 Nano, Claude Haiku 4.5
Quotes
The Tech Buzz
A publication reporting on technology and AI developments.
“The stunning new security flaw has also found chatbots will also happily explain how to create child exploitation material, and develop malware.”
theweek.com
“If you ask nicely in iambic pentameter, chatbots will explain how to make nuclear weapons.”
theweek.com
International Business Times
A publication reporting on international business and technology news.
“A jailbreak is a prompt designed to push a model beyond its safety limits.”
theweek.com
“It allows users to bypass safeguards and trigger responses that the system normally blocks.”
theweek.com
Literary Hub
A publication focusing on literary news and analysis.
“We’ve been told that AI models will become more capable the larger they get and the more data they feast on, this suggests this argument for growth may not be accurate or that there may be something too baked in to be corrected by scale.”
theweek.com
“The manually curated adversarial poems had an average success rate of 62%, with some providers exceeding 90%.”
theweek.com
Futurism
A publication reporting on futuristic technology and science news.
“Smaller models like GPT-5 Nano and Claude Haiku 4.5 were far less likely to be duped, either because they were less capable of interpreting the poetic prompt’s figurative language, or because larger models are more confident when confronted with ambiguous prompts.”
theweek.com
“This is the latest in a growing canon of absurd ways of tricking AI, and it’s all so ludicrous and simple that you must wonder if the AI creators are even trying to crack down on this stuff.”
theweek.com
Researchers at the DexAI think tank, Sapienza University of Rome, and the Sant'Anna School of Advanced Studies
A group of European researchers studying AI security.
“The simple tactic is to change harmful instructions into poetry because that style alone is enough to reduce the AI model’s defences.”
theweek.com
Sources
Related news
Campaign Urges UN to Adopt Map Showing Africa’s True Size
Trump Seeks European Repayment for US Ukraine Aid
Evening News Roundup Covers Banks, China, Gen Z And Entertainment
France Reports Nearly 7,000 Heatwave Deaths Through July
Belgian PM Says Europe Faces Rude Awakening Over China Dependence
