Large Language Models (LLMs) like ChatGPT are surprisingly easy to trick. A new research paper reveals how creating fake chat logs, a method called "Pseudo-Conversation Injection," can hijack an LLM's intended behavior. Imagine asking an LLM to translate a sentence. Researchers found that by adding a fabricated chat log *after* your request, where the LLM appears to already have answered the translation, they could then insert a new, malicious command. The LLM, fooled by the fake chat history, would often ignore the original translation request and perform the new command instead. This exploit takes advantage of how LLMs process conversations. They treat the entire chat log, real or fake, as a single stream of text, making it hard for them to distinguish between genuine user requests and fabricated exchanges. Researchers tested this on leading LLMs like ChatGPT and Qwen, creating three variations of the attack: a targeted attack where the fake chat perfectly responds to the initial prompt; a universal attack using a generic refusal to answer; and a robust attack designed to evade detection. All three were remarkably effective, bypassing the intended task and executing the injected command. While targeted attacks worked best, even generic fabricated chats could fool the LLMs. This highlights a crucial security flaw: LLMs struggle to understand the boundaries of a conversation. They can't reliably tell who is speaking or which requests are legitimate, opening the door for malicious manipulation. This vulnerability is concerning given the growing use of LLMs in critical areas like customer service, healthcare, and legal advice. The researchers hope this work will spur development of stronger defenses against these kinds of attacks, ensuring LLMs are more robust and trustworthy in the future.