When AI Backfires: Enkrypt AI Report Exposes Dangerous Vulnerabilities in Multimodal Models