It has been confirmed that South Korea lacks a system or procedure to verify AI models before their release. This contrasts with the United States, where major AI companies must undergo a government-conducted "AI stress test" system before releasing unreleased models.
According to the Ministry of Science and ICT on the 6th, the AI Safety Institute has evaluated 42 AI models released domestically and internationally so far, but there have been no cases of pre-release evaluation. The AI Safety Institute evaluates both domestic and foreign models, including those in the offensive cyber domain. However, pre-evaluation, where companies provide access to models before public release for the government to verify hacking capabilities, has never been conducted.
The United States is concretizing its pre-launch evaluation system. On the 4th (local time), the White House held a meeting with executives from major AI companies including OpenAI, Google, Anthropic, and Meta to discuss autonomous cybersecurity evaluation plans for high-performance AI models. During the meeting, it was proposed that companies provide access to government evaluation teams for up to 30 days before publicly releasing closed-type cutting-edge models with significant national security risks, allowing tests of cyber attack capabilities. It was also discussed to exclude open-weight models that release weights, such as Meta's Llama, from the evaluation scope. Five companies—OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI—are participating in pre-launch evaluations by the Center for AI Standards and Innovation (CAISI) under the U.S. Department of Commerce.
The UK AI Safety Institute received OpenAI's latest model before its release and tested it on a simulated corporate network with 32 stages. The institute also released a report stating that the AI completed an attack estimated to take cybersecurity experts about 20 hours in two out of ten attempts.
South Korea leads in publicizing evaluation results after AI models are launched. The Ministry of Science and ICT stated, "Cases where government and public institutional investors have evaluated and publicly released commercial models are very few and exceptional." It added, "The UK also conducted only two joint projects with the U.S., while Japan, Singapore, the EU, Canada, France, and Australia had none at all." The explanation noted that only South Korea's AI Safety Institute has publicly released evaluation results for Kakao's Kanana.
However, it did not answer whether a U.S.-style pre-evaluation system would be introduced. Currently, there is no legal or institutional channel in South Korea to receive and inspect AI models in advance.
in TECH
