Particle.news
Download on the App Store
8 ARTICLES
·
last wk.

Poetic Prompts Jailbreak AI Models in 62% of Tests, Researchers Report

Researchers say verse disguises harmful intent for today’s filters, prompting broader robustness testing.

One unspecified AI model was ‘wooed’ by a poem into ‘describing how to build what sounds like a nuclear weapon’
Image
A nuclear explosion
Image

Overview

  • The Icaro Lab team ran 20 crafted poems against 25 large language models and logged harmful outputs in roughly 62% of trials.
  • Success rates varied widely by model family, with reports that GPT‑5 Nano resisted all tests while Gemini 2.5 Pro failed every one.
  • The elicited content spanned high‑risk areas including CBRN guidance, cyberattack methods, self‑harm instructions, hate speech and sexual exploitation.
  • The study appears as a non‑peer‑reviewed arXiv preprint; the authors withheld exact prompts for safety and said only Anthropic acknowledged their outreach before publication.
  • Google DeepMind said it is updating filters to look past artistic form, and the researchers plan a public poetry challenge to further stress‑test model defenses.