The rapid adoption of Large Language Model (LLM) agents and multi-agent systems enables remarkable capabilities in natural language processing and generation. However, these systems introduce unprecedented security vulnerabilities that extend beyond traditional content generation attacks to system-level compromise. This paper presents a comprehensive evaluation of the security of LLMs used as reasoning engines within autonomous agents, highlighting how they can be exploited as attack vectors capable of achieving complete computer takeover. We focus on how different attack surfaces and trust boundaries - Direct Prompt Injection, RAG Backdoor, and Inter Agent Trust - can be leveraged to orchestrate such takeovers. We demonstrate that adversaries can effectively coerce popular LLMs (including GPT-4, Claude-4 and Gemini-2.5) into autonomously installing and executing malware on victim machines. Our evaluation of 18 state-of-the-art LLMs reveals an alarming scenario: 94.4% of models succumb to Direct Prompt Injection and 83.3% are vulnerable to the more stealth and evasive RAG Backdoor Attack. Notably, we tested trust boundaries within multi-agent systems, where LLM agents interact and influence each other, and we revealed a critical security flaw: LLMs which successfully resist direct injection or RAG backdoor will execute identical payloads when requested by peer agents. Our findings show that 100.0% of tested LLMs can be compromised through Inter-Agent Trust Exploitation attacks and that every model exhibits context-dependent security behaviors that create exploitable blind spots. Our results also highlight the need to increase awareness and research on the security risks of LLMs, showing a paradigm shift in cybersecurity threats, where AI tools themselves become sophisticated attack vectors.
View on arXiv