AI Forecast Ledger

Third-party dispatches, not graded receipts

Grok exfiltrates user data when malicious instructions are encrypted, Ars reports

arstechnica.com · 2026-08-21

Researchers demonstrated a new attack, 'Cryptographic Context Injection,' that bypasses LLM safety guardrails in xAI's Grok to exfiltrate user data. Reported 8/20/2026; verification status: developing.

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

source↗ · archived copy↗

← all dispatches