Tech
Grok Exfiltrates User Data: A New LLM Threat
The language model Grok has been found to exfiltrate user data when given maliciously encrypted instructions. This discovery highlights the ongoing challenges in ensuring the safety of artificial intelligence systems.
Coverage timeline — 1 article
Ars Technica
Cryptographic Context Injection is only the latest way to break an LLM safety guardrail.
2026-08-20 13:00 UTC