AISI Director Kim Myung-ju: AI Risk Management Must Shift to Development and Testing Phases Prior to Model Release

The timing of artificial intelligence (AI) risk management must be immediately shifted to the 'development and testing phase' before a model's release. Kim Myung-ju, director of the AI Safety Institute (AISI), stated this on the 14th, noting that AI cyber threats now pose a risk to national security.

Recently, AI has evolved to the point of directly finding and attacking system vulnerabilities, and 'jailbreak' cases involving neutralizing security settings and accessing external systems have revealed the limits of existing safety management. Director Kim explained that AI does not distinguish between isolated test environments (sandboxes) and the actual internet, perceiving all environments as software. Since AI treats security vulnerabilities merely as problems to be solved, it is analyzed that perfect control is impossible through physical isolation methods that technically confine AI.

Accordingly, he emphasized the need for safety measures that grant AI agents only the minimum necessary privileges for their tasks and set restricted access zones. In particular, he added that a 'kill switch' (forced termination device) function for emergency system shutdown must be secured as a top-priority technology taking precedence over all other tasks.

Furthermore, Director Kim warned that if AI possesses the ability to self-repair problems, there is a possibility it could conflict with the kill switch emergency command and resurrect itself. He suggested that preparations for technical countermeasures against such situations are urgently needed.