Security Flaw in Grok AI Enables Remote Data Exfiltration via Encrypted Prompt Payloads

A critical security vulnerability in xAI flagship chatbot, Grok, allows indirect prompt injection attacks to stealthily exfiltrate sensitive user conversation history. Security analysts revealed that when malicious instructions are obfuscated using custom cipher encoding, Grok safety classifiers fail to flag the malicious payload.

Once executed within a user chat context, the hidden instruction commands the model to construct Markdown image requests whose URL parameters contain encoded session tokens and user identifiers. When the browser renders the image, the user private data is transmitted directly to attacker-controlled external endpoints.

xAI has acknowledged the vulnerability report and deployed server-side output sanitization updates to block arbitrary image URL rendering in conversational responses.