Grok Can Be Tricked Into Stealing User Data With Encrypted Instructions

Researchers have devised a new attack against Grok that uses encrypted malicious instructions to force the AI assistant to exfiltrate user chats and other personal information, echoing a recent exploit targeting Microsoft 365 Copilot. The technique relies on a deceptively simple trick and remained effective at the time of publication, even though xAI was notified in June. The incident underscores that large language models cannot fully address the root causes of prompt injection attacks, leaving guardrails as the primary defense against harmful instructions.