New RouteHijack attack exploits MoE LLM vulnerabilities

Researchers have developed a new attack method called RouteHijack that targets Mixture-of-Experts (MoE) Large Language Models (LLMs). This attack exploits the routing mechanism within MoE architectures, identifying and manipulating safety-critical experts to bypass alignment safeguards. RouteHijack demonstrated a significant success rate across various MoE models, including vision-language models, highlighting a fundamental vulnerability in sparse expert architectures. AI

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

IMPACT Exposes a fundamental vulnerability in MoE architectures, necessitating new defense strategies beyond output-level alignment.

RANK_REASON Academic paper detailing a novel attack method on MoE LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• RouteHijack
• LLMs
• MoE
• arXiv

• paper
• safety