OpenAI and Korea AI Safety Institute sign high-risk AI evaluation agreement

The partnership will cover cybersecurity testing, Korean-language evaluation, and work on internationally applicable AI safety benchmarks.

OpenAI and the Korea AI Safety Institute will cooperate on high-risk AI evaluation, including cybersecurity and international benchmarking

OpenAI and the Korea Artificial Intelligence Safety Institute have signed a memorandum of understanding to cooperate on the evaluation of advanced AI systems in high-risk areas, including cybersecurity.

The agreement was signed at the Seoul office of the Electronics and Telecommunications Research Institute. It is the fourth agreement OpenAI has reached with a national AI safety institute, following arrangements in the United States, the United Kingdom, and Japan.

The partnership will cover safety evaluation methods, benchmarking, and the exchange of technical information. The Korea Artificial Intelligence Safety Institute and OpenAI will also examine how assessments can account for the Korean language and South Korea’s social context.

The agreement extends AI cooperation established through a separate memorandum signed by South Korea’s Ministry of Science and ICT and OpenAI in October 2025.

Working-level discussions will now determine the specific research projects, evaluation tasks, and timetable covered by the partnership.

Cybersecurity included in high-risk testing

The Korea Artificial Intelligence Safety Institute and OpenAI will exchange information on the evaluation of AI systems used in high-risk domains.

Cybersecurity is named as one of the initial areas of cooperation. The partners will consider domain-specific risks, evaluation methods, and examples of current practice.

The work will also cover high-performance AI models and autonomous AI agents, including how their capabilities and potential risks can be assessed before or during deployment.

Jiyeon Cho, Senior Researcher in AI Safety Policy and Strategic Cooperation at the Korea Artificial Intelligence Safety Institute, wrote on LinkedIn: "This MoU marks the beginning of collaboration on AI safety evaluation methodologies and best practices, including in high-risk domains such as cybersecurity."

The agreement does not commit the parties to a specific assessment product, public benchmark, or implementation deadline. Those details will be decided through subsequent technical discussions.

Evaluations will include Korean language and context

The partnership will examine how internationally used AI safety tests can account for language and local conditions.

The Korea Artificial Intelligence Safety Institute and OpenAI plan to exchange technical information relating to evaluation systems that reflect Korean-language use and South Korea’s social context.

That work could include the design of test data, benchmark tasks, and evaluation criteria intended to identify risks that may not appear in English-language assessments. The announcement does not specify which models or services will be tested.

The partners will also review challenges involved in creating AI safety frameworks that can be applied across different countries.

Cho wrote: "As AI safety is a shared global challenge, partnerships across institutions, sectors, and countries will be critical to advancing robust and trustworthy evaluation practices."

Agreement expands international safety cooperation

The memorandum adds South Korea to OpenAI’s network of formal relationships with national AI safety institutes.

The Korea Artificial Intelligence Safety Institute said the agreement would support international cooperation on the testing and verification of advanced AI models. OpenAI will contribute information from its model development, safety, and evaluation work.

South Korea’s Ministry of Science and ICT said the partnership builds on discussions held during two meetings this year between its Second Vice Minister and senior OpenAI representatives.

The Korea Artificial Intelligence Safety Institute and OpenAI will now begin working-level negotiations to confirm the first evaluation projects and schedule.