Naver Cloud Speeds Up Development of Security-Specific AI... “700B Model to Be Unveiled in July Next Year”

Naver Cloud is accelerating the development of a cybersecurity-specialized AI foundation model against the backdrop of the suspension of overseas model use due to U.S. export controls and a sharp rise in AI-based hacking. The model under development is a cybersecurity-specialized artificial intelligence foundation model, and Naver Cloud is betting on its own security ecosystem.

To that end, Naver Cloud has self-deployed 3000 NVIDIA B200 GPUs. It has also formed a consortium with a total of 32 participating institutions, and has a track record of being selected as an operator for the special-purpose model project. The target date is July next year, and the target scale is a 700B model.

Kim Jin-hui, an executive vice president at Naver Cloud, disclosed the development status and execution strategy for the special-purpose model at 'FISCON 2026' held on the 12th at the Conrad Seoul in Yeouido. Based on these progress updates and strategy, Naver Cloud is speeding up development of a cybersecurity-specialized AI foundation model.

Naver Cloud had prepared an AI security strategy based on overseas state-of-the-art models such as OpenAI's 'GPT' and Anthropic's 'Claude,' but around 20 days of inability to use those models occurred due to the fallout from U.S. export controls. Kim said that while establishing AI-based attack and defense strategies, the project came close to being halted after model use was suspended by U.S. export restrictions. He cited 'AI security sovereignty' and 'limits to real-world use' as the background for developing an in-house security model.

Amid that sense of crisis, Kim said the company formed its own team and planned to begin development of a security-specialized AI. He added that it entered the project when the special-purpose model call for proposals was announced.

Naver Cloud determined that it would be difficult to implement a global-level model with only 256 B200 GPUs supported by the government. It therefore decided to self-deploy 3000 B200 GPUs. Kim said this amounts to about KRW 200 billion when converted into 10 months of usage costs.

The special-purpose model began development on September 18 based on Naver Cloud's 'HyperCLOVA X (HCX)-Cyber.' The model has completed pretraining, and is now undergoing supervised fine-tuning (SFT) and reinforcement learning from verifiable rewards (RLVR). SFT refers to additional training that adapts a pretrained AI model to user needs, while RLVR is a learning process that rewards the model only when its outputs meet predetermined correct-answer criteria.

Naver Cloud plans to evaluate the performance of the first model with 300B parameters in February next year. It then plans to unveil a 700B model by July next year. The benchmark for the 700B model will be global models such as Z.AI's 'GLM-5.3.'

The criterion for technological differentiation is set around real-world field usability rather than benchmark scores. Kim explained that when deployed in the field, how quickly it can generate exploits and accurately identify vulnerabilities is more important than simply solving benchmark problems well. He also said recent research shows that AI's offensive and defensive capabilities are determined more by post-training than by pretraining, and that the company's own model is also focusing on that post-training process.

Naver Cloud is refining its deployment strategy for closed-network environments in the financial sector and public institutions. To support the use of the special-purpose model in sectors such as finance, power, defense, nuclear power, science and technology, telecommunications, and public administration, Naver Cloud is considering deployment methods such as on-premises operation without internet connection, SaaS, and API integration.

Naver Cloud is cooperating with the domestic security industry. The consortium has expanded from an initial 32 companies to 44 participating institutions, and Naver Cloud is conducting integration tests for security products from consortium members based on its own developed API and SDK. It also plans to promote support for embedding the HCX-Cyber model's core engine into domestic solutions.

Naver Cloud is strengthening cooperation with the financial sector, since the financial industry is widely regarded as the top target for hacking. The Korea Financial Security Institute has joined the consortium, and Naver Cloud plans to provide the special-purpose model to the institute so it can conduct a preemptive proof of concept (PoC). It also plans to provide participating financial companies with free credits for a certain period and dispatch FDEs, who will work to jointly resolve technical difficulties on site.

Kim pointed out that in the past, it took more than 2 years from vulnerability disclosure to an actual attack, but now the timeframe has shortened to as little as a day, and said this change is pushing security personnel increasingly toward a breaking point. He added that if high-quality data accumulated in the financial sector can be shared, it would be possible to deploy smarter AI at a lower cost than U.S. frontier AI.

Source: IT DAILY · Kim Ho-jun
Original: https://www.itdaily.kr/news/articleView.html?idxno=242886

References

• https://www.itdaily.kr/news/articleView.html?idxno=242886