Prompt Injection Attacks Threaten Multi-Agent Robotic Systems

A new study from arXiv (2608.00747v1) evaluates the vulnerability of LLM-based multi-agent robotic systems to prompt injection attacks. The research demonstrates that such attacks can induce adversarial actions and reduce task completion, with risks amplified in multi-agent settings through cross-agent contamination. Direct injections into task instructions and indirect injections via perception modules were tested across varying attack-goal complexities and injection strategies. Findings reveal that attacks can propagate between agents through shared prompt structures, with impacts depending on prompt composition and the targeted agent. The study underscores the need for robust security measures in autonomous systems that rely on large language models for planning and control.