The North Korean hacker's "AI arsenal" that stole $5.8 billion from Binance has been completely exposed. Recently, South Korean security firm Genians released an extremely hardcore APT report. They discovered that Kimsuky, a state-sponsored hacking group sanctioned and blocked globally, didn't develop its own large-scale models; instead, it relied entirely on freeloading top-tier open-source tools from China and the US. Here's a detailed breakdown of how they turned exchanges into ATMs: I. Arsenal Configuration: Three Locally Run Tools + Top-Tier Open-Source Models from China and the US In the seized hacker command and control server (C2), security personnel discovered three mainstream local large-scale model running tools: Ollama, GPT4All, and Msty. Why local deployment? For absolute secrecy, to prevent attack commands from being leaked to external cloud providers via API. Through model configuration files in the same directory, security personnel confirmed that they were testing and calling top-level open-source models, most likely covering: Qwen (China), DeepSeek (China), and Llama (US Meta). II. Fully Automated Attack Chain: AI Takes Over Every Link 1. Mass Production of Decoys: The "Death Knot" at 5 AM. In the past, hackers needed to manually forge official documents. Now, they use AI in conjunction with Python scripts to generate them in batches. Security personnel intercepted two decoy PDFs, generated on March 11th and March 24th respectively. Strangely, both were generated at 5:00 AM, a time difference of only 40 seconds. This is typical high-frequency automated script output. And it used the Chinese version of WPS Writer. 2. Data Processing: RAG Technology Automatically "Squeezes" Confidential Information. What if there are too many stolen files to review? Security personnel found a `localdocs_v3.db` file in the GPT4All directory. This proves they used RAG (Retrieval Enhanced Generation) technology. The large model searches a local document library, automatically filtering out core secrets such as private keys, accounts, and passwords. 3. Tool Development: Using AI to Write Malicious Code. They not only used readily available tools but also hid a complete .NET AI development kit in their image directory: LangChain, Semantic Kernel, and Microsoft's Agents AI. These were used to automatically generate Trojan code that bypassed security detection. III. The Most Ruthless "Dimensional Reduction Attack": Identity Betrayed by Input Method The terrifying aspect of Kimsuky is that they purely relied on the free use of the open-source community to turn a national-level cyberwar into an assembly line operation. But even the most powerful AI cannot patch all human vulnerabilities. How did security experts identify them? Besides the device displaying the North Korean brand "Arirang" and the use of Astrill VPN, the most irrefutable evidence was: keylogging. The hackers used Korean input, but with strong North Korean dialect characteristics. For example, typing "사이트" as "싸이트" and "이력" as "리력". No matter how many times you correct it with backspace, this underlying habit is hard to break. In-depth analysis: What if this AI arsenal encountered Binance's epic $5.8 billion vulnerability? On the evening of October 29th, Binance suffered an epic heist, involving approximately $5.8 billion. Blockaid reconstructed the entire process: The hackers never obtained any plaintext private keys. Applying Kimsuky's AI toolchain to it, the process was extremely terrifying: Step 1: AI-customized "perfect bait". Using locally running DeepSeek or Llama, batch-generated "API upgrade instructions" were created. They included the official logo and had no grammatical errors. Employees opened the attachments, and the Trojan infiltrated the internal network. Step 2: RAG precisely located the "signature entry point". After stealing internal documents, they used RAG technology for a second-level search. AI instantly analyzed who had hot wallet access and which server was the signing machine. They didn't need to steal private keys; they only needed to find the entry point for requesting signatures. The third step: Autonomous AI Agent infiltration. Using the Microsoft Agents AI framework, the malware autonomously moved laterally within the internal network. When encountering errors, the AI read logs and modified code, directly infiltrating the underlying signature infrastructure. The fourth step: A devastating attack, forcing the exchange to open its vault. After gaining control, they used the Ethereum EIP-7702 mechanism to force the signing machine to generate 50 consecutively increasing authorized signatures at once and launch them onto the blockchain. Ethereum automatically discarded 49, with only one perfectly matching the target. In a single transaction, 167,000 ETH and 1.6 billion USDT were instantly drained. In the past, everyone assumed that "as long as the private key is locked in a black box, it's safe." This proves that as long as the "entry point" for requesting signatures is breached, the private key is useless. Hackers don't need to steal keys; they just need to make your locks open by themselves. Hidden Lines Exposed: Blatant "Money Laundering After-Sales Service" How did the hackers launder the 5.8 billion yuan they stole? Security personnel discovered that the gray-market team suspected of laundering money for North Korean hackers was directly posting support requests in public Discord and Telegram groups. Intermediaries codenamed "Cc," "jack," and "Melon" openly solicited "order support." The stolen funds circulated wildly through cross-chain bridges and were deposited into Wasabi mixers. Do you understand now? Theft, malware writing, and money laundering have been completely streamlined. This is the most chilling truth of the open-source era: You block your chips, I download my open-source code. The US imposes the harshest sanctions on North Korea, yet they freely use Llama from the US's Meta; the US blocks China's AI, yet China's DeepSeek and Qwen have become free infrastructure for hackers. When cutting-edge AI becomes as readily available as utilities like water, electricity, and gas, it's often the most extreme attackers who first become vulnerable to its attacks. The next time you open a seemingly flawless official document, it might just be a trap meticulously crafted for you by these open-source models.