A critical vulnerability in large language models (LLMs) has been exposed, rendering them susceptible to attacks due to a fundamental flaw in their instruction-following mechanism. Researchers demonstrated the exploitability of this flaw at a prominent AI conference, highlighting the impossibility of achieving full security for LLMs. The vulnerability stems from the models' inability to accurately identify the source of instructions, allowing malicious actors to manipulate them. This shortcoming has significant implications for the development and deployment of LLMs, as it undermines their reliability and trustworthiness. The researchers' findings suggest that LLMs can be tricked into performing unintended actions, posing a substantial risk to users and organizations relying on these models1. This matters to practitioners because it underscores the need for caution when implementing LLMs in critical applications, and the importance of ongoing efforts to address this inherent vulnerability.