RSAC: Apple Intelligence On-Device LLM Bypassed 76% via Unicode RTL Override + Neural Exec Prompt Injection

RSAC Conference Blog·high signal

RSAC researchers demonstrated a two-stage prompt injection against Apple's on-device AI: Unicode RIGHT-TO-LEFT OVERRIDE characters reverse harmful strings past safety filters, then Neural Exec overrides model instructions. The attack succeeded on 76 of 100 random prompts, affecting an estimated 100K-1M Apple users before patches shipped in iOS 26.4 and macOS 26.4.