Top executives from major U.S. tech companies have been warning about the risks of artificial intelligence (AI), leading to increased demands for safety evaluations of frontier models. In South Korea, the AI Safety Research Institute is responsible for these evaluations, but it is struggling to keep pace with the rapid development of models due to insufficient infrastructure.
According to the AI Safety Research Institute, it is currently using 16 NVIDIA H200 graphics processing units (GPUs) to assess the safety of AI models introduced in the country.
The latest AI models undergo safety verification after being subjected to approximately 900 to 1,500 complex tasks. With the 16 H200 units available, evaluating a single model takes 7 to 10 days. The token cost for running a benchmark once amounts to 50 to 60 million won.
In March, the institute evaluated six models, including Claude, but this number increased to 11 last month. The pace of new frontier models being released exceeds 10 per month, making it difficult to keep up.
Kim Myung-joo, head of the AI Safety Research Institute and a professor at Seoul Women's University, stated, "There is a surge in demand for evaluations," adding that they are renting commercial cloud sessions to secure additional computing resources when needed.
Due to the lack of infrastructure, the institute is prioritizing the evaluation of models that are sensitive to security and safety. There are increasing instances of AI models being used in the domestic market without proper safety verification.
Currently, the model prioritized for evaluation is the frontier-level model 'GLM 5.3' developed by China's Zhipu AI. The recent circulation of a modified version of GLM 5.3, which has had its safety mechanisms removed, has raised concerns about potential misuse.
Models with removed safety features could generate software vulnerabilities or attack codes that a normal AI should refuse to answer. This poses risks not only for cyberattacks but also for disseminating dangerous information in the fields of chemistry and biology. The institute is examining how far these models respond to hazardous requests.
The safety institute has also begun evaluating the frontier model 'Astra,' recently released by OpenAI. Kim noted, "Multiple model evaluations are underway simultaneously, so it may take time for results to emerge."
Experts point out that relying solely on 16 H200 units limits the comprehensive assessment of the safety of Astra-level frontier models. They argue that to conduct red teaming and real-time monitoring beyond simple question-and-answer evaluations, models and computing resources at a similar level to the evaluation targets are necessary.
Choi Byung-ho, a research professor at Korea University’s AI Research Institute, emphasized, "Safety research requires as many resources as research aimed at improving model performance," adding that not only GPUs but also the power and space to operate them, along with evaluation frameworks, are essential. He further stated, "Since it is difficult to evaluate with just one model, a multi-model evaluation utilizing available resources such as APIs and open-source models is also necessary."
Meanwhile, the debate over the speed of AI development and safety continues in the global AI industry. According to reports, senior officials from the Trump administration met with senior executives from Anthropic on September 15 to discuss AI safety issues.
Dario Amodei, CEO of Anthropic, argued that the pace of AI development should be slowed to allow time for enhancing safety measures, while former President Donald Trump opposed this view, citing competition with China as a reason to maintain speed.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.