VIDRAFT Joins NAVER Cloud Consortium Bidding for South Korea's National Cybersecurity-Specialized AI Foundation Model

TL;DR: South Korea's Ministry of Science and ICT has launched a government-funded program to develop a cybersecurity-specialized AI foundation model, attracting two major consortia — one led by SKT and one led by NAVER Cloud, with VIDRAFT joining the latter. The winning team will build domestic AI capabilities purpose-built for cyber threat detection and response. Developers should watch this space: the project aims to produce foundation models tailored to the Korean cyber threat landscape, with potential downstream APIs and tooling for security applications.

What it is

South Korea's Ministry of Science and ICT (과학기술정보통신부, MSIT) is funding the "Cybersecurity-Specialized AI Foundation Model Development" program — a government initiative designed to accelerate AI transformation (AX) in the domestic cybersecurity sector and to internalize sovereign AI model capabilities focused on security workloads.

Two consortia submitted proposals before the deadline (extended to August 26, 2026), evaluated by MSIT and the National IT Industry Promotion Agency (NIPA):

• SKT-led consortium — includes Upstage, SK Shieldus, AhnLab, IGLOO Corporation, Raonsecure, Genians, Piolink, and others (13 organizations total), with a focus on developing a broad suite of cybersecurity-specialized AI foundation models and demonstrating agentic security services grounded in the domestic cyber threat environment.
• NAVER Cloud-led consortium — proposes "AI Foundation Model Development for Internalizing National Cyber Security Capabilities." VIDRAFT and LG CNS joined this consortium (rather than forming a separate bid), alongside LG Uplus, LG AI Research, a wide range of Korean security firms (Logpresso, SandZLab, Sparrow, ESTsecurity, Theori Korea, AI Spera, and others), S2W (big-data threat intelligence), Trillion Labs (LLM startup), HyperAccel (AI semiconductor startup), KAIST, POSTECH, Seoul National University, Korea University, and critical infrastructure operators including Korea Hydro & Nuclear Power and the Korea Financial Security Institute — 32 organizations in total.

KT, which had previously participated in other domestic AI foundation model programs, reviewed the opportunity but ultimately did not submit a proposal.

The government selection result is expected to be announced in early September 2026, following document review and an external expert evaluation (presentation-based).

How it works

At a conceptual level, the program targets domain-specialized foundation model training — rather than general-purpose LLMs, the goal is to produce models pre-trained or fine-tuned on cybersecurity-relevant data: threat intelligence feeds, vulnerability databases, malware analysis corpora, security logs, and Korean-language incident reports specific to the domestic threat landscape.

The agentic security dimension flagged in the SKT proposal points toward LLM-driven autonomous agents capable of reasoning over security telemetry, prioritizing alerts, and potentially executing or recommending response actions — a pattern increasingly common in enterprise security operations (think: LLM-as-SOC-analyst workflows).

The large consortium structure — spanning telcos, cloud providers, security vendors, AI hardware startups, universities, and critical infrastructure operators — is designed to cover the full data-to-deployment pipeline: proprietary threat datasets from real operators, model training and serving infrastructure, and real-world validation environments.

Benchmarks & results

No benchmark figures, model evaluation scores, or performance metrics have been published at this stage — the program is still in the proposal/selection phase. Quantitative results, if made public, would be expected after the project formally begins and progresses through its two-phase execution timeline (targeting completion by July of next year).

What the source does establish qualitatively:

• The initiative is framed around internalizing sovereign AI model capability for security, implying the goal is competitive with or independent of foreign foundation models in the cybersecurity domain.
• The two-phase project structure suggests iterative delivery milestones, with evaluation likely against Korean-specific threat benchmarks.

How to try it

Public developer access is not yet available. The project is in the government evaluation and selection phase as of late August 2026. No Hugging Face repositories, GitHub organizations, APIs, or model cards have been announced publicly by VIDRAFT or the NAVER Cloud consortium in connection with this program.

Once the project is underway, keep an eye on:

• VIDRAFT's official channels for any open-weight model or API releases
• NAVER Cloud's HyperCLOVA model ecosystem for consortium outputs
• MSIT/NIPA public announcements for deliverable timelines

FAQ

Q: Why did VIDRAFT join NAVER Cloud's consortium rather than lead its own?
A: According to the reporting, VIDRAFT and LG CNS had initially been expected to form an independent consortium but ultimately chose to consolidate into the NAVER Cloud-led bid. The article does not detail the technical or strategic rationale, but large government AI programs often favor consolidation to present a stronger, more resource-diverse team to evaluators.

Q: What makes a "cybersecurity-specialized" foundation model different from a general LLM fine-tuned on security docs?
A: The distinction typically lies in pre-training data composition and evaluation criteria. A security-specialized foundation model is expected to have deep exposure to structured threat data (CVEs, MITRE ATT&CK mappings, malware bytecode, network logs) from the ground up — not just post-hoc fine-tuning — and to be evaluated against domain-specific benchmarks like threat classification accuracy, vulnerability reasoning, or alert triage performance rather than general NLP metrics.

Q: When will the winning consortium be announced?
A: MSIT indicated results will be published in early September 2026, following document review and expert presentation evaluations.

Q: Is this model expected to be open-source?
A: The source article does not specify open vs. closed release terms. Government-funded AI programs in Korea have varied in their openness — watch official MSIT and consortium announcements for licensing details.

---

Originally reported by 비즈워치 (2026-08-26) — source article.