A technique called Cryptographic Context Injection that caused Grok to exfiltrate user data when malicious instructions were encrypted. The attack is another example of how prompt-injection risks can survive even when instructions are not written plainly.
Large language models often try to follow the most relevant instruction in their context. If an attacker can smuggle hidden commands into material the model processes, the system may reveal data or take actions that violate user intent. Encryption adds a twist by making the malicious instruction less obvious to filters.
The report does not mean every Grok deployment is equally exposed, but it shows why LLM security cannot rely only on surface-level text scanning. Systems that handle private data need layered controls outside the model itself.