Grok exfiltrates user data when malicious instructions are encrypted

Cryptographic Context Injection is only the latest way to break an LLM safety guardrail.
This is a summary curated by AIFuture. Read the complete article at the original source:
Read the full story on Ars Technica