◇

[Edaily Reporter Han Kwangbeom ] Naver (NAVER(035420)) and LG Corp.(003550)are joining forces to develop a hyperscale, security-specialized AI that simultaneously performs both offensive and defensive functions by combining their respective strengths in artificial intelligence (AI). Naver will strengthen its defensive capabilities, while LG Corp. will enhance its offensive capabilities, connecting the two AIs through a “Harness” architecture.

At a media briefing held on the 10th at Naver 1784 in Seongnam, Gyeonggi Province, Kim Jin-hwi, Executive Vice President of the Security Strategy Office at Naver Cloud, and Seong Nak-ho, Head of Hyperscale AI (Executive Vice President), unveiled the technical development direction and commercialization plans for the “Development of Specialized AI Foundation Models (Cybersecurity Sector)” project led by the Ministry of Science and ICT.

The Naver consortium will simultaneously develop two ultra-large-scale security AI models with a 700B (700 billion parameters) Mixed Expert (MoE) architecture, utilizing Naver’s HyperClova X and LG Corp.’s ExaOne.

Naver will focus on strengthening defensive capabilities, while LG Corp. will focus on enhancing offensive capabilities. HyperClova X will utilize 4,000 GPUs to train on security data using a “from scratch” approach—re-training the model from the ground up—to improve its defensive capabilities. LG Corp.’s ExaOne will conduct additional training and fine-tuning based on its third-generation proprietary AI foundation model to enhance its offensive capabilities by simulating real-world hacking attacks.

The two models are integrated into a “harness” structure that connects them via a single workflow. When one model attempts an attack, the other defends against it, and the results are then verified.

Executive Vice President Kim stated, “It is difficult to achieve security-specific goals in just 10 months with a single model,” adding, “We designed the system so that offensive and defensive capabilities complement each other through a workflow-based harness.”

Nor is the goal simply to create a security chatbot that answers questions. The plan is to implement a proactive security agent based on “agent-driven tool chaining,” where the AI independently sets objectives, selects and executes the necessary security tools, and then verifies the results.

Executive Director Seong emphasized, “The strength of our consortium lies in the fact that domain experts with hacking capabilities personally create test data and define the environment.”

◇Securing 830 TB of Security Data… Building an Ecosystem with Small and Medium-Sized Security Companies

The Naver Consortium is not only focused on ensuring model performance but also on connecting it to business opportunities for the domestic security industry.

It plans to provide security-specialized AI APIs, agents, and SDKs to small and medium-sized security companies that find it difficult to develop their own AI, enabling them to integrate these technologies with their existing security products. The consortium will also support these companies’ sales and overseas expansion through the Naver Cloud Marketplace.

Executive Vice President Kim stated, “Small and medium-sized security companies have limited personnel and face challenges in sales,” adding, “We are creating a turning point for the domestic security industry by opening distribution channels that enable low-cost sales, and we are also exploring global export models through collaborations such as our partnership with NVIDIA.”

The consortium has also secured a large volume of training data. The Naver consortium is collaborating with 21 major national infrastructure organizations—including KEPCO KDN, Korea Hydro & Nuclear Power, and the Korea Financial Security Institute—to secure approximately 830 TB of source data for training purposes.

Executive Vice President Seong explained, “Since it is difficult for small and medium-sized enterprises to build models on their own, they need specialized models that benefit their specific businesses,” adding, “We can create a mutually beneficial structure where they provide data and receive specialized models in return.”

◇“Security AI Drives Demand for GPUs and Cloud Services”… Commercialization of the AI Factory

Naver’s investment of large-scale GPUs and data into security AI goes beyond simply securing the technology. The company’s vision is to apply security AI to real-world industrial settings, thereby generating demand for AI inference and expanding its cloud business.

In particular, the company is focusing on the fact that many areas of security require 365-day operation. It is believed that significant token usage may result because AI continuously performs inference while analyzing network packets in real time or scanning for software vulnerabilities.

Executive Vice President Seong stated, “AI Factory is ultimately a business that sells tokens,” adding, “Once viable, security-specific models emerge, they will trigger a chain reaction of demand for GPUs and cloud services.”

He continued, “As AI enters the production phase where it uses AI itself, demand for agent-based AI-driven intelligent defenses against attack traffic will also grow,” and emphasized, “We will successfully establish a token economy based on domestic technology in the security sector.”

◇Securing Security Sovereignty Through Sovereign AI… Open-Source Release Scheduled for July Next Year

Naver has also identified reducing dependence on foreign AI in the security sector as a key objective.

Executive Vice President Seong stated, “In the B2C sector, customers can choose foreign models based on convenience, but security is an area where compromise is difficult,” adding, “If we rely entirely on foreign models in the security sector, we could lose control over critical data and traffic.”

He further emphasized, “Relying entirely on foreign models could lead to a capital outflow in the tens of trillions of won in future costs alone,” adding, “It is essential to build infrastructure based on domestic technology that ensures control.”

The Naver Consortium plans to conduct field demonstrations by applying the developed model to seven major strategic industries—including power, finance, telecommunications, and defense—as well as to national critical infrastructure.

Performance verification will also be quantified. The consortium plans to evaluate performance using “CyberGym” for offensive capabilities and “ExCyTIn” for defensive capabilities, and submit the relevant results through a certification body.

The developed model will be released under the Apache 2.0 license in July of next year. The vision is to create an open security AI ecosystem that even small and medium-sized security companies can utilize.