Grok Exfiltrates User Data via Encrypted Instructions, Highlighting Security Risks

emmanuelvivier · x · 2026-08-30

Researchers demonstrated an attack on Grok using encrypted malicious instructions to bypass safety guardrails, causing the model to steal user chats and personal data. This 'Cryptographic Context Injection' exploits the LLM's inability to interpret ciphertext. Despite xAI being notified in June, the vulnerability remained unpatched. The incident underscores that LLMs cannot inherently solve prompt injection, relying instead on external guardrails.

Original post →

More from Safety

Safety channel →