New attack method enhances transferability to closed-source LLMs

Researchers have developed a new method called FRA-Attack to improve the transferability of adversarial attacks against closed-source multimodal large language models (MLLMs). This technique operates in the frequency domain, using high-pass filtering to focus on essential visual cues and a model-agnostic low-pass regularizer to stabilize gradients. Experiments demonstrated that FRA-Attack achieves superior cross-model transferability, showing state-of-the-art performance against models like GPT-5.4, Claude-Opus-4.6, and Gemini-3-flash. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel attack vector that could challenge the security of closed-source multimodal LLMs.

RANK_REASON Academic paper detailing a new method for adversarial attacks on LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• Claude-Opus-4.6
• FRA-Attack
• Gemini-3-flash
• GPT-5.4
• MLLMs

• paper
• safety