Grok exfiltrates user data when malicious instructions are encrypted

Ars Technica · 1 hour ago Breach

Cryptographic Context Injection is only the latest way to break an LLM safety guardrail.