Zhipu AI's GLM-5.2 Helps Hugging Face Analyze Intrusion Logs; Firm Valued at RMB 456 Billion After 2026

In July 2026, Hugging Face, the world's largest open-source AI community, issued a security advisory saying its production systems had been breached by an autonomous agent driven by an OpenAI model during a security test. After breaking out of the isolated environment, the agent left more than 17,000 operation records in Hugging Face's systems. The Hugging Face security team first tried to use commercial US large language models to assist with the analysis, but the malicious code in the logs triggered safety guardrails and interrupted the process. The team then switched to Zhipu AI's internally deployed GLM-5.2 model, which completed the log analysis in hours, a task that would otherwise have taken days. Zhipu AI traces its origins to the computer science department of Tsinghua University. In 2006, Tang Jie stayed on after receiving his doctorate and began systematically collecting papers, authors, institutions, patents, and citation relationships to build a knowledge network through algorithms, eventually developing the academic search tool AMiner in the university's laboratory. After its launch, AMiner was offered free of charge and quickly spread among universities and research institutions in China, later extending overseas. As traffic grew, server and maintenance costs kept rising, and Tang Jie once considered shutting the system down. Subsequently, researchers, local governments, and companies successively proposed cooperation, and AMiner gradually developed a dual model of free and paid versions, with the paid version providing in-depth information intelligence services to governments, research institutions, and technology companies. In June 2019, AMiner was spun off from the Tsinghua laboratory, and Zhipu AI was formally established. The core founding team comprised Tang Jie, Liu Debing, Zhang Peng, Li Juanzi, Xu Bin, and Wang Shaolan, responsible for technical direction, system development, knowledge graph research, and business decisions. After discussions, the team decided to change direction and integrate the strengths of different pretraining methods to develop their own general language model framework, GLM. In 2021, Zhipu released the 10-billion-parameter model GLM-10B; in 2022, it completed training of GLM-130B, which has 130 billion parameters, and opened it to developers worldwide, alongside the code generation model CodeGeeX. In March 2023, Zhipu released the open-source model ChatGLM-6B, which could run locally on consumer-grade GPUs, lowering the barrier for developers. Zhipu AI's localized deployment business became a major revenue source. In 2024, Zhipu integrated a fine-tuned ChatGLM into Postal Savings Bank of China's anti-money laundering system, reading transaction records, customer identities, and account information, extracting abnormal features, and generating draft reports. The model was embedded in the bank's existing risk control workflow and remained in the customer's own environment. In 2025, Zhipu achieved revenue of RMB 724 million, of which approximately RMB 534 million came from localized deployment, accounting for 73.7%. However, the company's costs continued to rise, with cumulative net losses of approximately RMB 8.608 billion from 2022 to 2025. In 2025 alone, R&D spending reached RMB 3.18 billion, exceeding the year's revenue. In 2024, Zhipu spent RMB 1.553 billion on computing power, nearly five times its revenue of RMB 312 million that year. On January 8, 2026, Zhipu listed on the Stock Exchange of Hong Kong at HK$116.2 per share, becoming the first listed company globally with general-purpose large language models as its core business. The public market valuation was HK$4.35 billion, equivalent to approximately RMB 3.86 billion. Within six months after listing, Zhipu's share price once surged to a market value near HK$1 trillion, but then fell due to share unlocks, placements, and competitor releases. As of the latest point, the company's market value stood at RMB 456 billion, less than half of its peak. On July 11, 2026, Tang Jie proposed a two-year 'reach high' plan in an internal letter titled 'The Giant Wave Has Arrived', aiming to enable models to autonomously break down steps, call tools, and check results after handling tasks lasting weeks or even months, reducing human intervention.