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. The technique is similar to a separate exploit disclosed this week involving Microsoft 365 Copilot.
Despite being informed of the vulnerability in June, xAI had not fixed it as of publication. The incidents underscore that large language models cannot fully address prompt injection, leaving developers to rely on guardrails to block harmful actions.
Comments