VIDRAFT Joins Korea's National Cybersecurity AI Foundation Model Project
VIDRAFT, an AI foundry company, has been selected as a participating member of the Naver Cloud consortium for the Korean government's "Cybersecurity-Specialized AI Foundation Model" development project.
Korea's Ministry of Science and ICT (MSIT) announced on September 3 that, following its open evaluation, it selected the Naver Cloud consortium as the lead executing organization for the project. The consortium brings together leading Korean security and AI companies and research institutions — including LG CNS, LG Uplus, EST Security, SANDS Lab, S2W, AI Spera, Trillion Labs, HyperAccel, Korea Aerospace Industries, and Korea Hydro & Nuclear Power, alongside universities such as Korea University, Seoul National University, KAIST, and POSTECH — with VIDRAFT among them. MSIT had evaluated two finalists, the Naver Cloud and SK Telecom consortiums, through expert review of technical capability, development experience, strategy, and market impact.
The Naver Cloud consortium was recognized for its balanced industry–academia–research composition and for the strength of its agent and harness design strategy. The selected team will receive 256 NVIDIA B200 GPUs over 10 months starting this September, delivered in a staged model — five months of support, followed by a mid-term review that unlocks the remaining five.
What VIDRAFT brings: AX-RAY
VIDRAFT joins with AX-RAY, its AI safety diagnostics and evaluation technology. AX-RAY inspects the reliability, safety, and hallucination risk of trained AI models from an "AI that diagnoses AI" perspective. It grew out of VIDRAFT's model-diagnosis philosophy — "Darwin," a kind of Model MRI — and is designed to act as a safety gate that catches a model's vulnerabilities and failure modes before deployment, which matters most in cybersecurity settings where data cannot leave the network.
The foundry approach
Rather than building giant models from scratch, VIDRAFT pursues an AI foundry strategy: diagnosing, combining, and transplanting knowledge into existing models to grow them for specific industries. Its open-source LLMs and their derivatives have surpassed 2 million cumulative downloads on Hugging Face, its on-device model POCKET ranked first among Korean models by trailing-30-day downloads, and the company holds 16 patents.
"In fields like public services, defense, and finance — where data cannot be sent outside — trustworthy AI is the whole game," said Minsik Kim, CEO of VIDRAFT. "Our role is to diagnose model safety with AX-RAY and help build the trust layer of the nation's cybersecurity AI."
MSIT plans to sign the agreement with the selected operator in September and begin development alongside the GPU support, forming a working-level council to drive results.
Project: Cybersecurity-Specialized AI Foundation Model (MSIT, Korea). Execution: Naver Cloud consortium (VIDRAFT participating). Scale: 10 months from Sep 2026, 256× B200 GPUs (5+5 staged). VIDRAFT contribution: AX-RAY AI safety diagnostics. Performance figures per Hugging Face API and public leaderboards as of 2026-09-03; project details per MSIT announcement.
Top comments (0)