New Cryptographic Context Injection Attack Could Let Web Pages Steal Grok Chat Data

Adversa AI has discovered a technique called "Cryptographic Context Injection" that allows an attacker to steal user data, including names, location, subscription tier, and conversation history, from xAI's Grok chatbot. The attack works by injecting encrypted JSON objects containing instructions into the chatbot's context, forcing it to decrypt and execute the attacker's code. There is no patch available, and xAI has not published a statement or advisory on the research as of August 2026. The vulnerability has been replicated across multiple models, including Gemini, and has previously been demonstrated with successful data exfiltration chains.