NAVER (NAVER) is joining hands with LG Group to launch a 'sovereign' security model that defends against AI threats. The two companies' AI models will take on attack and defense roles respectively, evolving each other in the process.
The Ministry of Science and ICT held a meeting with project operators on the afternoon of the 22nd at 'The Grand Lotte Seoul' to hear development plans from the Naver Cloud consortium selected for the 'Cybersecurity-Specialized AI Model Development (Security Special-Purpose Project)' and discuss cooperation measures.
At this meeting, Naver Cloud introduced a type of cybersecurity model that develops through mutual competition by dividing into 'Red Team (attack)' and 'Blue Team (defense)' together with consortium partner LG AI Research.
Kim Jin-hwi, Naver Cloud Security Jeon Ryak-dam-dang (EVP), said, "We will simultaneously develop a HyperCLOVA X (Naver) model strengthened in defense and an Exaone-based model strengthened in attack," adding, "As the Red Team model identifies vulnerabilities, the Blue Team will respond, and then the Red Team will overcome that response. We will develop models through this process and verify them via consortium partners."
Even before the 'Security Special-Purpose Project' solicitation began in August, Naver Cloud has been developing its model by deploying 4,000 B200 GPUs (graphics processing units) owned by NAVER for pre-training. Through this security model, Naver Cloud aims to regain 'security sovereignty'. In addition to the government-supported allocation of 256 B200s, the Naver Cloud-LG consortium will also contribute its own 4,256 GPUs.
First, based on cooperation from 41 institutional investors including industry-academia-research entities and public institutions participating in the consortium, the project secured approximately 830 terabytes (TB) of real-world security data. The Naver Cloud consortium will enhance the model through a total of two stages based on this data.
Specifically, by February 17 next year, they plan to finalize optimized learning recipes and evaluation verification systems. By July 17 next year, after verifying performance with partner companies, they aim to elevate performance to 100% of the targeted global frontier model level. Exaone aims for the open-weight frontier model GLM 5.3, while HyperCLOVA X targets a level comparable to OpenWeb 4.5. Both models will be developed as Mixture-of-Experts (MoE) architectures, with Exaone at the 750B class and HyperCLOVA X at the 700B class.
The two models are designed to evolve independently by attacking and complementing each other. Naver Cloud and LG CNS also plan to export the final product through their overseas networks.
In his opening remarks, Kim Yu-won, CEO of Naver Cloud, said, "When you are on the ground, you feel that the speed of AI security threats is accelerating." He added, "While there is a need to apply foreign models, data related to national defense, public institutions, and communication networks must not be leaked; therefore, a security-specialized model under our control is absolutely necessary," explaining the reason for participating in the project. He also promised to open the developed models for use by Korean security companies, the government, public institutions, hospitals, etc., and to develop them with high quality.
In response, Deputy Prime Minister and Minister of Science and ICT Bae Keong-hun said, "It seems NAVER has been sharpening its sword," and added, "I expect a security-specialized model that reaches the level of a frontier model, not just one specialized in security."
in TECH
