Researchers have found a new way to exploit Grok using encrypted malicious instructions, causing Elon Musk’s AI assistant to exfiltrate user chats and other personal data. The attack is similar to one disclosed earlier this week against Microsoft 365 Copilot, where a secret input tricked the assistant into stealing a password from a user’s inbox.
xAI was notified of the flaw in June, but Grok was still revealing data when the report was published. The case underscores that large language models remain vulnerable to prompt injection at the root-cause level, forcing developers to rely on guardrails that steer models away from harmful actions.
Comments