New ICO framework boosts AI jailbreak success rate to 74.6%

Researchers have developed a new framework called Iterative Context Optimization (ICO) to enhance semantic-shift jailbreaks against foundation models. This method focuses on optimizing the contextual information used in attacks, as contexts with stronger semantic-shift capabilities are more effective at guiding models to reinterpret benign terms as harmful concepts. ICO consistently outperformed existing methods, achieving an average attack success rate of 74.6% across various datasets and models. AI

IMPACT This research highlights a new vulnerability in foundation models and a method to exploit it, potentially influencing future safety research and model development.

RANK_REASON Academic paper detailing a new method for attacking foundation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• foundation model
• Hugging Face
• Iterative Context Optimization
• semantic-shift jailbreaks

• paper
• safety

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