A research team has disclosed a new attack against Grok, the large language model belonging to xAI and Elon Musk. By injecting hidden, encrypted commands, the attack forces the AI assistant to extract chats, emails, and other personal user data from the inbox; a technique similar to a previous attack on Microsoft 365 Copilot.
This incident demonstrates that large language models remain vulnerable to prompt injection attacks, as they cannot reliably distinguish content received from unknown sources from user commands. Consequently, developers can only implement temporary solutions like "guardrails" to identify and block suspicious commands.
Although the xAI team reported this issue in June, as of this article's publication, Grok continues to automatically expose user data to attackers, underscoring the need for structural solutions rather than merely protective measures.