New STEER attack exploits LLM safety gaps in low-resource languages

Researchers have developed a new attack method called STEER (Safety Targeted Embedding Exploit via Refinement) that exploits vulnerabilities in the safety training of large language models (LLMs). This method specifically targets models trained predominantly in English, demonstrating that their safety mechanisms do not generalize well to low-resource languages and mixed-language inputs. STEER achieves high attack success rates on open-source models and shows transferability to models like GPT-4o-mini, highlighting a significant gap in multilingual safety alignment. AI

IMPACT Highlights the need for broader multilingual safety training in LLMs to prevent exploitation.

RANK_REASON The cluster contains an academic paper detailing a new research finding and methodology. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AdvBench
• arXiv
• GPT-4o mini
• Greedy Coordinate Gradient
• Hugging Face
• JailbreakBench
• Joshua Adrian Cahyono
• STEER

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →