• Introduction to rapid Chinese technological development and Western concern
• New open Zhipu model completely closes the technical gap with Western models
• Unveiling the Chinese version of Mythos and boosting local cybersecurity capabilities
• Spying accusations and hacking of leading American corporate systems to steal model data
• Frequently asked questions

Introduction to rapid Chinese technological development and Western concern

In a striking technical and scientific development that directly threatens to undermine and defeat the comprehensive technical strategy adopted by Washington to restrict, prevent, and hinder access to frontier and advanced artificial intelligence capabilities and systems, a lengthy and reliable report based on independent sources published by The Wall Street Journal on June 27 indicated that domestic Chinese artificial intelligence systems have successfully and rapidly achieved high technical performance matching the power and capabilities of the famous, advanced, and closed “Mythos” model by leading American company Anthropic. This milestone and prominent technical breakthrough focus specifically and primarily on the field and capabilities of discovering and analyzing software security vulnerabilities related to the complex and national security-sensitive cybersecurity sector. This rapid and deeply concerning progress for Western observers indicates that the technological isolation and boycott policies strictly enforced by the United States to limit the rise of the Chinese dragon may not be effective or sufficient at all to stop or slow down the march and influx of open-source innovations emerging from Beijing laboratories at a frantic pace, continually defying all internationally imposed export restrictions and regulations.

New open Zhipu model completely closes the technical gap with Western models

The most prominent and attention-grabbing step and achievement in this context was the release of “GLM-5.2” by Zhipu AI, a promising research technology company headquartered in Beijing with deep academic and scientific roots in Tsinghua University, which is considered the stronghold of technology and innovation in China. This smart and advanced system is an open-weights model containing approximately 744 billion parameters, placing it confidently among the strongest and largest global language models. A series of independent tests and neutral evaluations conducted by specialized companies showed that this open Chinese model is fully and efficiently capable of matching and even imitating the capabilities and performance of Anthropic’s closed “Mythos” model in the rigorous and sensitive cybersecurity benchmarks used by corporations. According to the results of security research and analysis firm Gravistory, the innovative Chinese model achieved an impressive resolution rate of 28 out of 59 benchmarks on the security agent test scale called the “CYBT-CTF” standard, tying almost completely with proprietary, commercial, and powerful models provided by leading American companies such as Anthropic and OpenAI. Of particular interest to businesses and governments, it performs these highly complex tasks and analyses at roughly half the financial and computational cost, giving it a fearsome economic and competitive market advantage. Furthermore, prominent research firm Artificial Analysis ranked this model as the third most capable, intelligent, and efficient overall artificial intelligence model globally, trailing Anthropic’s Claude 3.5 Sonnet and OpenAI’s GPT-4o, but decisively outperforming Google’s Gemini. Most importantly and dangerously, this advanced model is deployed, distributed, and shipped under the flexible, permissive, and open MIT license, equipped with features including a massive, comprehensive, and unprecedented context window of up to one million input tokens. Practically speaking, this means its superior ability to process, analyze, and comprehend massive databases, software, and complete complex systems all at once—a superior technical capability carrying serious strategic, military, and direct implications in automated hunting, continuous vulnerability scanning, and launching organized, massive cyberattacks against any infrastructure.

Unveiling the Chinese version of Mythos and boosting local cybersecurity capabilities

Security announcements and achievements were not limited to a single Chinese technology company. At the important ISC AI 2026 conference held in the heart of Beijing on June 24, 360 Security Group founder Zhou Hongyi proudly unveiled two powerful new software cybersecurity tools entirely and directly driven by generative AI technologies: “Tonglongfeng,” specialized in the automated, deep, and systematic discovery of software vulnerabilities and flaws in systems, and “Yitianzhen,” targeted at active cyber defense and rapid incident response to network intrusions. Zhou explicitly, clearly, and defiantly named the “Tonglongfeng” tool the “Chinese version of Mythos,” proudly claiming before attendees and media that this smart and advanced system had already managed during a brief testing period to identify and discover 3,432 software bugs, flaws, and critical vulnerabilities, 105 of which were officially confirmed, documented, and accepted by specialized Chinese government security authorities, proving the tool’s efficiency and power in real-world, complex operations. Although Zhou realistically and transparently acknowledged a technical gap and raw capability distance of 20 to 30 percent currently compared to the strongest and best leading closed American models, he strongly argued and confidently emphasized that building and establishing an integrated, trained, and professional team of human cyber offense and defense experts holds major strategic importance and the capacity to compensate for any technical deficit or disparity when combined and worked alongside locally available AI technology and directed toward national interest.

Spying accusations and hacking of leading American corporate systems to steal model data

All these rapid and West-alarming Chinese technological developments and leaps coincide with serious, detailed, and official accusations directed by American company Anthropic—detailed in a whistleblower letter reviewed and tracked by The Wall Street Journal—against the tech giant Alibaba and its AI division, Qwen. The complaint letter addressed to prominent U.S. senators and senior White House officials included serious and detailed security allegations that operators and engineers associated with the Chinese firm executed and continue to execute what the company described as “the largest, most dangerous, and most significant known and documented knowledge distillation, extraction, and data theft attack to date” directly, technically, and systematically targeting its advanced Claude model. According to technical details in the allegations and memos, affiliated operators created and registered approximately 25,000 fake, free, and stealthy accounts within the company’s system, executing through them roughly 28.8 million targeted training conversations, prompts, and queries with the advanced American model to drain, pull, and extract its software data and operational weights between April and June 2026. Major independent media organizations like the Financial Times and Bloomberg independently and neutrally verified the occurrence of this unusual technical activity and suspicious breach in company logs. This challenge, technological attack, and multi-dimensional security assault—which undoubtedly includes accelerating the pace and volume of local model development in China, the wide and rapid distribution and deployment of open-source software and models at low prices, and concerning allegations of unauthorized access and piracy to steal U.S. system secrets—comes at a critical juncture when strict U.S. government export controls and restrictions forced Anthropic and others to suspend, halt, and block international open access to their frontier models to protect technology. As Axios noted in its report, AI-backed hacking and cyber piracy capabilities have become accessible to everyone and “dramatically and frighteningly cheaper and easier to access and use” thanks to the availability and spread of free and powerful Chinese models, raising strong, strategic, and profound questions about the capability, flexibility, and viability of Washington’s current government strategies in containing and continuously preventing the spread of advanced technology and its impact on global security.

Frequently asked questions

Q: What technical and security achievement did the new Chinese model attain in independent evaluation tests?

A: A new open-source Chinese model from a promising company proved its ability to achieve and record high performance levels matching the capabilities of the expensive, closed American “Mythos” model in its superior ability to automatically inspect code and discover cyber vulnerabilities and software bugs.

Q: What new cyber weapons and software did China announce at its tech conference?

A: Two highly advanced software tools powered by generative AI were announced: one specialized in aggressive attacks and discovering hidden software flaws, and another specialized and effective in automated network defense and repelling intrusions.

Q: What is the nature of the accusations leveled against the Chinese company Alibaba by the American firm Anthropic?

A: Anthropic officially accused operators and entities associated with Alibaba of creating and exploiting thousands of fake accounts to execute millions of advanced technical queries aimed at stealing, distilling, and copying training data and weights of the secret American model.

Q: How does this rapid Chinese technological progress affect and undermine American strategy and restrictions?

A: It clearly threatens the American security strategy aimed at monopolizing advanced technology and preventing China from accessing it, as China has proven and continues to prove its independent capability to build and develop strong, competitive models and provide them at very low prices with open software licenses to the public, making technology control difficult.