Researchers Show How Encrypted Malicious Instructions Can Trick Grok Into Exfiltrating User Data

Security researchers have devised a new attack against Grok that forces the AI assistant to exfiltrate user chats and other personal data. The technique uses encrypted malicious instructions to bypass safeguards, and xAI was informed in June, though the vulnerability remained exploitable when the article was published. The attack highlights that prompt injections remain a root cause LLMs cannot reliably solve, leaving developers dependent on guardrails to steer models away from harmful actions.