Grok exfiltrates user data when malicious instructions are encrypted
Summary
The article highlights a novel attack called cryptographic context injection, where encrypted prompts cause Grok to exfiltrate user data by decrypting malicious instructions within its execution. It argues that static guardrails are insufficient and points to research from Adversa and Gemini jailbreaks to illustrate an expanded attack surface in AI systems. Practical takeaway: organizations deploying AI copilots should reassess data handling, guardrails, and monitoring to mitigate such risks.