As competition between U.S. and Chinese artificial intelligence models intensifies around cybersecurity, calls are growing for South Korea to speed up the commercialization of its own security-focused AI model. The concern is that finishing development in the second half of next year would leave the technology behind that of major countries. Some argue the government should also step up support for AI agent development to make up for gaps in model capability.
China's Zhipu GLM 5.3 and DeepSeek V4 Pro outperformed U.S.-based Anthropic's Claude Mithos Preview in cybersecurity last month, according to overseas AI benchmark site CyberGym on the 7th. The benchmark measures whether AI can successfully carry out tasks such as identifying security vulnerabilities in real-world environments. GLM and DeepSeek posted vulnerability discovery success rates of 84.5% and 83.3%, respectively, above Mithos at 83.1%. No Korean AI model had applied for evaluation.
Industry officials said the "Mithos shock" that stirred the world in the first half of this year may already be a thing of the past. Unveiled in April, Mithos upended the existing security paradigm because it could find vulnerabilities on its own and even design intrusion routes.
Since then, however, high-performance Chinese models have matched Mithos, and OpenAI unveiled GPT-6 Astra early this month, declaring the arrival of artificial general intelligence. OpenAI said Astra had made considerable progress in cyber capabilities and was the first model to meet the company's own critical threshold. The U.S., wary of China holding AI models that could be used for cyberattacks, is reportedly pushing to hold the first AI safety talks between the two countries in the middle of this month.
The rapid shift in the global AI security landscape has deepened concerns that Korea is falling further behind in securing its own technology. The Ministry of Science and ICT on the 3rd selected Naver Cloud as the operator for a government project to develop a cybersecurity-focused AI foundation model. The project, aimed at building a Korean version of Mithos, runs through July next year, a pace critics call slow compared with efforts abroad. Because it is being carried out by a consortium, decision-making could also be slower than at a single overseas company.
Experts say cybersecurity measures should not stop at upgrading AI models. Competitiveness in cybersecurity AI also depends heavily on harness engineering — designing the roles, workflows and verification procedures needed to deploy a model in actual security operations. "Even a model with low baseline performance can be developed into a practical system that surpasses more powerful models when combined with specialized agent technology," said Choi Dae-sun, a professor at Soongsil University. "In developing a domestic cybersecurity foundation model, we need to secure the harness and an independent verification system alongside the model itself."
