[Edaily Reporter Shin Yeong-bin ] Naver (NAVER(035420)) has set its sights on developing artificial intelligence (AI) for cybersecurity. The company aims to release a security-specialized AI model that goes beyond the government’s expectations and approaches “frontier-level” capabilities.

On the 22nd, the Ministry of Science and ICT held a kickoff briefing for the “Development of Cybersecurity-Specialized AI Models” project at the Lotte Hotel in Jung-gu, Seoul, and unveiled the specific development plans of the Naver Cloud consortium.

Naver Cloud and the AI Research Institute at the LG Corp.(003550) plan to develop two types of 700B-class AI models—one specialized for attack and the other for defense—using 4,512 graphics processing units (GPUs) and approximately 830 terabytes (TB) of security data.

Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon said on the same day, “It seems Naver has been sharpening its sword,” adding, “I expect they will produce specialized models that are not just specialized, but on par with frontier-level models.”

Deputy Prime Minister Bae emphasized that as AI-enabled cyberattacks accelerate, South Korea must develop its own response capabilities.

He explained, “We believed it was important to quickly secure a certain level of frontier-class models as well as models specialized for cybersecurity,” adding, “We began with the idea that if we could develop cybersecurity-specialized models faster than anyone else, we might be able to resolve the various issues we currently face.”

Developing Both Offensive and Defensive AI Together

A total of 41 organizations—including 33 consortium members and 8 partner institutions—are participating in this project. Roles will be divided into data collection and processing, model development, evaluation and verification, and field demonstration and deployment, with the project proceeding from development through demonstration over the next 10 months.

The consortium will simultaneously develop two AI models specialized in attack and defense.

Naver Cloud’s HyperClova X will enhance “defensive” capabilities by detecting, analyzing, and responding to cyberattacks. LG Corp. AI Research’s K-ExaOne 2.0 will take on the “offensive” role of identifying vulnerabilities in applications and source code.

The two models will train against each other. When the attack model identifies vulnerabilities in source code and pathways that could lead to actual attacks, this information is used as training data for the defense model. Once the defense model devises a countermeasure, the attack model is trained to bypass it, thereby improving the performance of both models simultaneously.

The computing resources allocated to this effort are substantial. The government is providing 256 NVIDIA B200 GPUs, while Naver Cloud is contributing 4,000 of its own B200 GPUs. LG Corp. AI Research also plans to contribute 256 H200 GPUs. Combined, this amounts to 4,512 GPUs.

Naver Cloud has been developing a security-specialized HyperClova X using 4,000 B200 units since last August, employing a “from scratch” approach that involves training the model entirely from the beginning.

LG Corp. is conducting additional training based on the K-ExaOne 2.0 750B MoE (Mixed Expert) model to enhance its attack detection capabilities. Ultimately, the security-specialized models to be developed by both companies aim for a 700B-class MoE scale.

Approximately 830 TB of security data required for training has been secured as an initial step.

This will combine threat and operational data held by Naver Cloud and LG CNS (#LG CNS) with real-world data from national critical infrastructure and the security industry field provided by consortium participants. A data preprocessing system will also be established to automatically process the source data into a format suitable for AI training.

Kim Jin-hwi, Executive Vice President of Naver Cloud, stated, “The 41 organizations will work as one team to develop the models using proven, independently developed domestic technology,” adding, “We will immediately deploy 4,512 GPUs and train the models using real-world national-level data.”

Field Testing in Finance, Power, Semiconductors, and Defense… Model to Be Released as Open Source

The developed

model

will be deployed at actual national critical facilities and industrial sites. It will be tested to verify its ability to respond to real cyber threats in sectors such as finance, power, science and technology, telecommunications, semiconductors, defense, and aerospace. Naver Cloud’s clients and LG Corp. affiliates will also participate in the field testing.

The system will also integrate with existing products from domestic security firms. The plan is to create a suite of domestically developed AI security products that perform AI-based vulnerability assessments, respond to cyberattacks, and investigate security incidents through integration via APIs, SDKs, and Agent AI. The scope of application will be expanded to include closed networks, such as those in the public and defense sectors, that are isolated from the outside world.

The company plans to release the model as open source. It will provide vulnerability assessments to small and medium-sized enterprises (SMEs) that lack security personnel and budgets, and will also make the Naver Cloud Marketplace available so that domestic security firms can use the model to develop new AI security products.

Kim Yu-won, CEO of Naver Cloud, stated, “It is extremely difficult for critical information related to defense, the public sector, or telecommunications networks to be leaked overseas,” adding, “Since AI cannot function properly without sufficient data, we need security-specific models under our control.”

He continued, “I believe the results of this project should not be utilized in a way that allows only Naver Cloud or participating organizations to have exclusive access,” adding, “We will prepare to ensure it can be widely used by Korean security companies, as well as government and public institutions, schools, and hospitals.”

CEO Kim emphasized, “We have promised a very high level of quality,” adding, “This is not merely a slogan or a declaration; in 10 months, we will demonstrate our true capabilities through various benchmark scores and validations.”