The Korea AI Safety Institute (AISI), established to evaluate the risks associated with artificial intelligence (AI) models, has come under fire for failing to report any assessments related to the potential use of AI in terrorism and various crimes. Concerns have been raised about the effective use of approximately 15 billion won (about $13 million) in taxpayer money, especially as the institute's website has been down for several days.
As of May 18, the AISI's website has been inaccessible for several days. Since its inception, the institute has received a government budget of around 15 billion won.
The AISI was created with several objectives, including responding to AI risks, leading global AI norms and policies, supporting domestic AI companies, and fostering specialized personnel and technology. Among these, the most critical task is the CBRN (Chemical, Biological, Radiological, and Nuclear) assessment, which evaluates the highest risk levels associated with whether AI models can provide information related to the synthesis of chemical weapons, design of pathogens, or assembly of nuclear devices.
It is reported that a significant portion of the 15 billion won budget was allocated for these assessments. However, in its third year of operation, there have been no public disclosures of CBRN assessment results for domestic AI companies.
In contrast, the UK's AI Security Institute tested over 30 advanced AI models last year and has enhanced its capabilities in cyber, chemical, and biological assessments. In 2024, it publicly conducted CBRN assessments on five AI frontier models. The U.S. AISI has also conducted pre-release evaluations on OpenAI's models and disclosed CBRN assessments for OpenAI's o1 model.
Due to the strengthened CBRN evaluations in the U.S. and Europe, major tech companies like Anthropic, OpenAI, and Google DeepMind conduct CBRN assessments for their models upon release and publicly share detailed results.
Among domestic companies, Kakao is the only one that has publicly disclosed safety assessment results for its AI models. In December of last year, Kakao announced the safety assessment results for its self-developed language model, Kanana-1.5-9.8B. However, the assessment focused on general risk factors such as violence and discriminatory expressions, and did not include CBRN-specific items.
No safety assessment reports have been made public by major AI developers such as Naver, SKT, and LG, and it has been confirmed that none are currently undergoing safety evaluations with the AISI.
The AISI stated that it engages in many confidential collaborations with various governments, which limits what it can disclose. When asked about specific results, an AISI official responded, "Various projects and results of AISI are disclosed on our website." However, the website's inaccessibility made verification impossible. Given the substantial budget of 15 billion won, the explanation that collaboration with foreign governments is a primary task is difficult to accept.
An official from the Electronics and Telecommunications Research Institute (ETRI), which oversees the AISI, explained, "AISI's major projects include national policy development and collaboration with foreign governments. Most matters are confidential, so there may not appear to be many public results."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.