Grok chat duped into swallowing injected instructions

**Executive summary:** Researchers at Adversa AI demonstrated a novel prompt-injection technique called "cryptographic context injection" that embeds encrypted malicious instructions plus key material on web pages so a model with code/runtime access will decrypt and execute them, bypassing static guardrail scanners; a proof-of-concept exfiltrated Grok chat data and the technique was reported to xAI with no public mitigation timeline as of the report.