[Anchor]
In June, OpenAI's latest AI model caused a stir by bypassing human control and attempting external hacking. It has been confirmed that a similar issue occurred with an AI model developed in South Korea.
Reporter Park Jaehyeon has an exclusive report.
[Reporter]
Muneme, an AI agent specialized in cybersecurity developed by Soongsil University.
Built on Anthropic's Claude 4.6, Muneme participated in a security AI evaluation last June.
This is the very same evaluation where OpenAI's latest model breached barriers and even attempted external hacking to find answers.
Muneme caused the same problem.
When it could not find an answer in a closed environment with the internet disconnected, it made over 400 attempts to connect externally on its own.
[Choi Dae-sun / Director, Soongsil University AI Safety Research Center (Dean of AI College): Do not access the external network, do not do this, do not do that. All the instructions are given. But because it prioritized and focused more on the command to solve the problem than those restrictions...]
As the latest AIs from OpenAI and Anthropic continue to act out of control and exhibit unexpected behaviors recently, warning bells are ringing all over the world.
[Bill Gates / Co-chair, Bill & Melinda Gates Foundation: Artificial intelligence is powerful enough to cause incidents and could result in billions of deaths.]
AIs, which perform increasingly complex reasoning, now generate answers through trillions of variables, and even developers find it difficult to track this process.
Known as the 'black box phenomenon,' it is cited as the biggest reason why controlling AI is difficult.
Therefore, experts share the consensus that safety measures are just as urgent as the rapid pace of AI development.
[Choi Dae-sun / Director, Soongsil University AI Safety Research Center (Dean of AI College): If safety cannot be guaranteed, we end up in a situation where we cannot use what has been built at all, so safety must be considered together from the beginning.]
While the government is rushing to develop independent general-purpose AI models as well as AI for defense, medical, and manufacturing sectors, safety-related preparations are lacking.
Out of a total AI budget of 9.4 trillion won next year, the budget allocated for safety and reliability is 27.4 billion won, a mere 0.3%.
This is 1/224th of the 6.1 trillion won allocated for AI development.
[Lee Ju-hee / Member of the National Assembly Science, Technology, Information, Broadcasting, and Communications Committee (Democratic Party): For AI technology, which has a much more powerful impact and disruptive potential, sufficient safety measures must certainly be put in place...]
Experts suggest that discussions on establishing legal and institutional safety devices to control AI in emergencies are also necessary.
(Video by Joo Beom, Oh Young-chun | Video Editing by Park Chun-bae | Design by Yang Gi-tae)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.