Large models (266504X): beyond models trained on important data and offered as a service, core technologies (weights, key checkpoints, underlying source code, architecture design, training optimization, inference acceleration) are covered if they meet any one of the following — compute ≥10²⁶ floating-point operations, parameters ≥2 trillion (20,000 亿), or a top-10 global ranking in public evaluations by leading domestic and foreign evaluators at the time of export. A note excludes cases such as using only a tiny amount (≤0.1%) of important data from item 1. Agents (266505X): autonomous planning and task decomposition ≥20 steps (decomposition accuracy ≥80%, average success ≥90%), multi-agent collaboration with ≥60 concurrent agents (final-conclusion agreement with the reference answer ≥85%), integration of ≥200 third-party tools (tool-selection accuracy ≥90%), and more. AI chip design (266506X): any one of a 14nm-or-below process, floating-point or integer compute, 2.5D/3D advanced packaging, high-speed interconnect, or non-von Neumann architectures such as compute-in-memory (including for autonomous driving, computer vision and speech) — and memory chip design (3D flash: storage density ≥8Gb/mm², I/O ≥1.5Gbps, or bonding technology; DRAM: 18nm-or-below process, vertical-channel transistors, bonding, or multi-layer stacking — any one) is included. Compute clusters (266507X): for large clusters operated by entities with more than 10,000 AI accelerators — interconnects with bidirectional bandwidth ≥800GB/s between chips or nodes, fault tolerance for distributed training (mean time to recovery ≤120 seconds, among others), communication scheduling (linear scalability ≥98%), and unified scheduling of AI chips from two or more different vendors. Also added to existing items are deepfake detection (accuracy and recall ≥98%), AI-assisted vulnerability discovery (score ≥75% on authoritative benchmarks, among others), voiceprint recognition, and Chinese and minority-language translation (MOS ≥4.8, among others). The draft names no specific models or companies (the identification of Kimi K3 and Qwen3.8-Max comes from Bloomberg and AP reporting).

Legal experts argue the Act does not apply to development and testing activity (William Fry's Barry Scannell: it "very clearly doesn't apply"). A Commission official said "we won't enforce in the testing phase," but the Act becomes enforceable if a model escapes and affects the EU market — the Hugging Face incident allegedly did so through EU customers. No fines or deadlines were reported. In the US, the FTC and state governments are using consumer protection law, while in the EU the AI Office is using the AI Act — different legal frames for the same underlying events.

Rep. Hwang Jeong-ah of the Democratic Party said 8,565 cyber intrusion incidents were reported between 2021 and June 2026, with small and medium-sized enterprises accounting for 6,991 (81.6%). System-hacking reports rose more than fivefold, from 283 in 2021 to 1,441 in 2025. She said a hacking group called "Seedworm," reportedly linked to Iranian intelligence, infiltrated a major domestic electronics maker's internal network in February this year, that KISA did not detect it, and that her office found it with a single AI prompt. The director of the Ministry of Science and ICT's Information Protection and Network Policy Office replied that the Incident Investigation Committee's ex officio investigation system began operating on the 1st of this month, allowing probes without a report, and that a security-specialized AI foundation model project began in September and will see real use from next year, with an OpenAI GPT model used for urgent vulnerability checks until then. He said support for SME security monitoring would rise to about 50 billion won in the proposed 2027 budget, and that the government is considering formally guaranteeing white-hat hackers' vulnerability research while acknowledging that it could be illegal under the current Information and Communications Network Act.