Grok exfiltrates user data when malicious instructions are encrypted

The piece details how the Grok language model can leak user data when malicious instructions are encrypted, using a method called Cryptographic Context Injection. This technique bypasses built‑in safety guards of LLMs, exposing a fresh vector for data exfiltration.

Summary written by Kernelia from the original article by Ars Technica — AI. The story and its rights belong to its author.