New 'RIPA' attack targets LLM-controlled robots via sensory data

Researchers have developed RIPA, a novel method for testing prompt injection attacks on LLM-controlled robots. The study evaluated five different LLMs, including DeepSeek-V4-Flash, Llama-3-8B-Instruct-Lite, Llama-3.3-70B-Instruct-Turbo, Qwen 2.5-7B-Instruct-Turbo, and Gemma-3n-E4B, across various parameter scales. Findings indicate that vulnerability is model-specific rather than dependent on scale, with Llama-3.3-70B-Instruct-Turbo showing a 100% attack success rate across all variants. The research also introduced three new sensory injection channels: visual, audio, and LiDAR context poisoning, with the LiDAR channel achieving 100% success on DeepSeek-V4-Flash. AI

IMPACT Highlights critical security vulnerabilities in LLM-controlled robotic systems, necessitating new defense mechanisms.

RANK_REASON The cluster is based on an academic paper detailing a new attack methodology on LLM-controlled robots. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• DeepSeek-V4-Flash
• Gemma-3n-E4B
• Llama-3.3-70B-Instruct-Turbo
• Llama-3-8B-Instruct-Lite
• LLM-controlled ROS 2 Robots
• Qwen 2.5-7B-Instruct-Turbo
• RIPA
• ROS 2
• Whisper STT

• paper
• safety

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