AX-RAY 与 K-MYTHOS：韩国联盟打造的安全专用 AI 基础模型

AX-RAY & K-MYTHOS: Inside Korea's Consortium-Built Security-Specialized AI Foundation Model TL;DR: A 33-organization Korean consortium led by Naver Cloud is building K-MYTHOS ("K-미토스"), a security-domain foundation model that fuses threat intelligence from malware, dark web, network, and cloud sources into a dual-model attack/defense architecture. VIDRAFT contributes AI safety diagnostics and evaluation via its AX-RAY framework, which covers 117 inspection checklist items. Developers working in SecOps, threat intelligence, or AI safety should watch this project closely as it targets real-world deployment across Korean security products. What it is K-MYTHOS is a government-supported, security-specialized AI foundation model developed by a 33-institution consortium, with Naver Cloud as the lead organizer. The project has a dual-model design: An attack AI trained on the EXAONE base model (developed by LG AI Research) A defense AI trained on Naver Cloud's HyperCLOVA X base model The consortium spans security vendors, universities, and public institutions, each contributing domain-specific threat data and tooling. Target deployment surfaces include email security, network isolation (망연계), access control, and Security Operations Center (SOC) products — validated across seven industry verticals. VIDRAFT's role in this consortium is AI safety diagnosis and evaluation. Its AX-RAY tool performs structured safety assessments of AI systems across 117 inspection items, functioning as an independent auditing layer over the models produced by the broader consortium. How it works The architecture is best understood as a data fusion + dual-model pipeline: 1. Multi-source threat data aggregation Participating organizations contribute heterogeneous security datasets: Malware & endpoint telemetry — from consumer AV and EDR products Dark web & Telegram threat intelligence — ransomware groups, leaked credentials, attack TTPs IP/domain/asset exposure data — from internet-scanning platforms that map attacker infrastructure Network packet and IPS event data — for detecting attack behaviors at the wire level SIEM/SOAR log correlation — for incident timeline reconstruction Cloud-native security posture data (CNAPP) — misconfigurations, excessive permissions, runtime threats 2. Dual-model adversarial training environment Rather than training a single monolithic model, the consortium separates offensive reasoning (attack AI on EXAONE) from defensive reasoning (defense AI on HyperCLOVA X). This adversarial framing allows each side to be tested against realistic attack scenarios generated by the other. 3. AX-RAY safety evaluation layer (VIDRAFT) Before any model reaches a product integration stage, AX-RAY performs a structured 117-item inspection. This checklist-driven approach is conceptually similar to red-teaming frameworks like NIST AI RMF or OWASP LLM Top 10, but tailored to the Korean security AI context. It serves as a formal gating mechanism for AI safety and reliability in deployment. 4. Product integration & field validation Validated models are integrated into commercial security products (email gateways, network isolation appliances, SOC platforms) and tested in live industrial environments across seven verticals. Benchmarks & results The source article does not publish quantitative benchmark scores for K-MYTHOS or AX-RAY at this stage. What is reported qualitatively: AX-RAY covers 117 distinct inspection checklist items — the most concrete evaluation-scope figure disclosed The project targets deployment validation across 7 industry verticals (specific verticals not fully enumerated in this coverage) One consortium member (ESTsecurity) has already shipped a commercial AI-enhanced EDR product using a predecessor government foundation model program, suggesting the pipeline from consortium model → commercial product has prior proof points Expect quantitative results to be published as field validation completes. How to try it Public access is not available at this time. K-MYTHOS is an ongoing government-consortium R&D project. No public Hugging Face repository, GitHub release, or API endpoint has been announced for the consortium model or AX-RAY as of this reporting. If you want to track availability: Monitor Naver Cloud's official developer channels and HyperCLOVA X documentation Watch LG AI Research's EXAONE public repositories on Hugging Face for any security-domain variants Follow VIDRAFT's official announcements for any AX-RAY evaluation tooling releases FAQ Q: What exactly does VIDRAFT's AX-RAY evaluate — the training data, the model outputs, or both? A: Based on the source, AX-RAY is positioned as a safety diagnosis and evaluation tool for AI systems, operating via a structured 117-item checklist. The source does not break down what proportion of items target data provenance, model behavior, or deployment configuration — but the framing as a "security AI safety diagnostic" suggests it covers the full AI system lifecycle rather than a single artifact. Q: Is K-MYTHOS a fine-tuned model or a full foundation model trained from scratch? A: It is built on top of existing Korean foundation models — HyperCLOVA X for the defense AI and EXAONE for the attack AI — making it a domain-specialized derivative rather than a from-scratch pre-train. The security-specific training comes from the multi-source threat data contributed by consortium members. Q: How does the dual attack/defense model architecture improve over a single security model? A: The adversarial separation means the attack model can simulate realistic offensive behavior (malware, TTPs, evasion) that the defense model must learn to counter. This is conceptually similar to GAN-style adversarial training, but applied at the agent/model level using real-world threat data rather than synthetic noise. Originally reported by IT조선 (2026-09-14) — source article.