Researchers have investigated prompt injection attacks on robots controlled by Vision-Language Models (VLMs). The first study systematically examined physical prompt injection using adversarial text in the robot's visual field, categorizing attacks and testing them on three frontier VLMs (GPT-4o, Gemini 2.5 Flash, Qwen3-VL-32B). These attacks succeeded at rates between 5% and 29.4%, with authority-impersonating and negation attacks proving cross-model transferable. The second study focused on multi-agent robotic systems, demonstrating that prompt injection can induce adversarial actions and propagate between agents, impacting task completion and safety. Both studies highlight the vulnerability of VLM-controlled robots to such attacks and explore potential mitigation strategies. AI
IMPACT Highlights significant security vulnerabilities in AI-controlled robotics, necessitating the development of robust defenses against adversarial manipulation.
RANK_REASON The cluster consists of two academic papers published on arXiv detailing research into prompt injection attacks on robots controlled by Vision-Language Models.
Read on arXiv cs.MA (Multiagent) →
- arXiv
- Large language models
- LLM-based multi-agent robotic system
- Multi-agent robotic systems
- Prompt Injection Attacks
- Neha Nagaraja
- Gemini 2.5-Flash
- GPT-4o
- M. A. Viraj J. Muthugala
- Qwen3-VL-32B
- Robots
- Vision-Language Models
AI-generated summary · Google Gemini · from 4 sources. How we write summaries →
