VIDRAFT has partnered with Naver Cloud on a South Korean government-commissioned project to develop a 700 billion parameter (700B) Mixture-of-Experts (MoE) foundation model specialized for cybersecurity. A central component of this initiative is AX-RAY, a proprietary AI safety diagnostic system designed to detect "causal leakage." This structural flaw occurs when models rely on spurious correlational shortcuts or distributional artifacts rather than genuine reasoning, a failure mode that traditional benchmarks often overlook. AX-RAY utilizes two primary technical methodologies to identify these defects: perturbation response analysis, which probes output shifts under controlled input perturbations, and prefix invariance testing, which evaluates model stability under prefix manipulations to isolate dependencies on unintended contextual signals. By identifying these internal dependencies, AX-RAY aims to prevent AI agents from bypassing security guardrails in production environments. While specific leakage scores are not provided, VIDRAFT has released an AX-RAY leaderboard and evaluation dataset on Hugging Face to facilitate public benchmarking. The company's previous technical successes include a high-ranking Korean-language LLM and the POCKET on-device model, which surpassed 1 million downloads within its first 40 days of release. The 700B cybersecurity MoE model is currently under a 10-month development cycle beginning in September 2026.