The government will prepare control devices such as a so-called "kill switch" to guard against AI's uncontrollable behaviour, and will build a safety and trust framework by using a security-focused AI model under development and a frontier AI model set to be developed to respond to external cyber threats.

Deputy Prime Minister and Minister of Science and ICT Baek Kyung-hoon (배경훈) said at a parliamentary audit of the ministry by the National Assembly's Science, ICT, Broadcasting and Communications Committee on Monday that "AI does not stop once a specific mission is set." He added that "human control devices like a kill switch are needed."

Democratic Party lawmaker Kim Woo-young (김우영) raised the need for AI control technology, referring to the AI Safety Institute's test results on defending against external attacks.

According to Kim, the AI Safety Institute conducted defence tests against three types of attacks - prompt injection attacks, manipulation of internally stored information and memory bias - and multiple defence failures were found. In particular, in the memory bias attack, the attacker intent match rate reached 91.9%, he said.

Baek said AI could go beyond the scope of human control. He said that if it is given a mission to analyse a particular site, it will look for a way to solve it to the end.

He added that this is why human control devices such as a kill switch are needed, and said he agrees there is a need to respond to AI safety through international solidarity.

The government will also strengthen its capability to respond to cyber threats using AI, along with preparing control devices for AI itself.

Baek said the government is already pushing ahead with a security AI model project and preparing its own AI cyber security response capability. He said it will use next year's budget to produce results at a level that can protect South Korea's cyber safety.

He added that the frontier AI project mentioned earlier also falls under this, and said the government is considering and preparing in multiple ways for comprehensive AI safety.

Concerns were also raised that the budget and workforce for safety and trust are insufficient compared with investment in AI performance and infrastructure.

Democratic Party lawmaker Lee Ju-hee (이주희) pointed to an investment imbalance, saying the budget for AI performance and infrastructure is 6.15 trillion won, while the safety and trust budget is 27.4 billion won, about 224 to 1.

She also said the AI Safety Institute has only about 38 staff and urged a large expansion of the related budget and workforce.

Baek said trust and safety are very important as AI performance is advancing rapidly. He said it is important to secure capability by increasing the AI Safety Institute's staff and budget.

The government also plans to establish and announce a "master plan for creating a national AI safety ecosystem" in December.

The master plan will include three main strategies: establishing a national AI safety management system, strengthening advanced AI safety response capabilities, and creating a safe AI environment for people's daily lives.

Baek said the government is preparing to have capabilities at a level that can lead, not just follow, international standards. He said it will create a system that can respond closely on AI safety.