As cyber attacks suspected to involve artificial intelligence (AI) agents target the domestic financial sector, the national budget for evaluating AI risks has remained largely stagnant. The budget for the AI Safety Research Institute for the upcoming year is set to increase by 24% compared to this year, but most of the additional funding is focused on research and development (R&D), leaving the actual budget for evaluating AI models and responding to threat trends effectively unchanged.
On October 6, Rep. Lee Jeong-heon of the Democratic Party, a member of the National Assembly's Science, Technology, Information, Broadcasting and Communications Committee, revealed that the government budget proposal for the AI Safety Research Institute in 2027 is 15.944 billion won, a 23.8% increase from this year.
The overall budget increase is primarily driven by R&D. The R&D budget for developing AI safety technologies for trustworthy AI is set to rise from 3.975 billion won to 6.891 billion won, marking a 73.3% increase and accounting for about 95% of the total budget increase. The Ministry of Science and ICT plans to continue supporting existing projects on anti-deepfake detection and resilience against adversarial attacks while launching two new projects aimed at enhancing AI safety evaluation and verification technologies.
In contrast, the non-R&D budget for evaluating currently released AI models, collaborating with domestic and international safety organizations, and analyzing threat trends has only increased from 8.9 billion won to 9.053 billion won, a mere 1.7% rise. While funding for developing mid- to long-term evaluation technologies has expanded, the operational budget needed to meet the growing demand for immediate model evaluations has effectively been frozen.
The evaluation burden on the AI Safety Research Institute is rapidly increasing. The number of AI models evaluated by the institute rose from six in March to eleven in August, an increase of approximately 83%. Evaluating a single agent-type AI model typically takes 7 to 10 days, with 3 to 4 days required for cyber threat assessments. The cost of running a benchmark once can reach 50 to 60 million won.
The institute is also relying on external support for high-performance graphics processing units (GPUs). The evaluation office is currently using two B200 servers and 16 GPUs rented from a national AI project for large-scale and frontier-level AI model evaluations. There are no additional GPUs purchased by the office or confirmed for future acquisition. The institute plans to secure additional resources through rentals.
In response, Rep. Lee Jeong-heon criticized the situation, stating, "While long-term R&D budgets are increasing, neglecting the GPU infrastructure of the 'safety evaluation office' that must verify the immediate risks of emerging global frontier AI is a serious case of bureaucratic inertia that risks missing the critical window for national cybersecurity." He emphasized the need to expand dedicated evaluation personnel, operational budgets, and computational resources to not only respond to threats but also to proactively verify and control the attack capabilities of AI models.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.