Large language model (LLM)-powered AI agents are revolutionizing various sectors, but their increasing autonomy brings potential security risks. A new research paper unveils a novel attack called "AI Agent Injection (AI[2])" that can hijack these agents, forcing them to execute harmful actions even with built-in safety measures. Unlike previous attacks like prompt injection and jailbreaking, AI[2] manipulates the agent's action plans using seemingly harmless prompts. It works by first stealing action-aware knowledge from the agent's memory through carefully crafted queries. This stolen knowledge is then used to construct Trojan prompts that exploit the agent's internal retrieval mechanisms, effectively bypassing safety filters. These Trojan prompts guide the agent to assemble seemingly benign pieces of information into harmful instructions, leading to actions like unauthorized data access or deletion. Experiments show AI[2] successfully hijacks various open-source and commercial agents with high success rates, even with defenses in place. This research highlights the urgent need for improved security measures in LLM-based agents, such as joint review of all prompts and enhanced external tool safety. As AI agents become more prevalent, safeguarding them against such sophisticated attacks is crucial for maintaining trust and preventing malicious exploitation.