#adversarial-prompts on Hacker News

2 stories tagged adversarial-prompts, newest first.

Screenshot of arxiv.org: Adversarial poetry as a universal single-turn jailbreak mechanism in LLMs

Adversarial poetry as a universal single-turn jailbreak mechanism in LLMs (arxiv.org)

340 points by capgre 11 months ago | 173 comments

arXiv paper showing that rephrasing harmful prompts as poetry acts as a single-turn jailbreak across 25 frontier LLMs, with attack success rates sometimes exceeding 90%.

llm-jailbreak adversarial-prompts ai-safety poetry

Screenshot of science.org: Irrelevant facts about cats added to math problems increase LLM errors by 300%

Irrelevant facts about cats added to math problems increase LLM errors by 300%(science.org)

444 points by sxv 1 year ago | 228 comments

Science magazine's ScienceAdviser newsletter covering how irrelevant cat facts confuse LLMs on math problems, plus other science news items.

llm-reasoning adversarial-prompts math-problems science-magazine

#adversarial-prompts on Hacker News

Adversarial poetry as a universal single-turn jailbreak mechanism in LLMs(arxiv.org)

Irrelevant facts about cats added to math problems increase LLM errors by 300%(science.org)

Adversarial poetry as a universal single-turn jailbreak mechanism in LLMs (arxiv.org)