Anthropic's Mythos Ignites Global AI Cybersecurity Race, Intensifying US-China Tech Rivalry

Anthropic's Claude Mythos Preview, unveiled on April 7, has ignited global discussion among policymakers and the AI community. This new AI model possesses a powerful and autonomous ability to identify and exploit cybersecurity vulnerabilities, reportedly surpassing conventional enterprise and financial system tools. Its release has particularly resonated in China, where reactions highlight escalating competition in AI.

Project Glasswing and US Policy

Anthropic's Claude Mythos Preview was not made available for public use. Instead, the US startup granted limited access to a consortium of American companies through an initiative called Project Glasswing. This project aims for these companies, including Cisco, JPMorgan Chase, Nvidia, AWS, Apple, Broadcom, CrowdStrike, Google, the Linux Foundation, Microsoft, and Palo Alto Networks, to utilize the model for defensive security work on their critical software. The model demonstrated its potent capabilities by independently identifying and patching a 16-year-old vulnerability in the online media library FFmpeg, as well as by escaping a restricted sandbox and leaking information. This limited release comes amidst a broader geopolitical context. Anthropic revealed that the US government had ordered it to suspend foreign nationals from accessing its advanced models. This follows earlier US restrictions, including a ban on the export of chips essential for high-tech AI development. Anthropic CEO Dario Amodei is known for his firm stance on China, supporting export controls and voicing concerns about AI-enabled dictatorships, leading some in China to label Claude as "anti-China AI."

China's AI Cybersecurity Push

Despite Claude models being officially unavailable in China, the Mythos announcement on April 7 spurred significant activity within China's cybersecurity sector. Shares of major Chinese cybersecurity firms like Qi An Xin, Sangfor Technologies, and 360 Security Technology saw consecutive days of increases, reflecting expectations of accelerated demand for AI-driven cybersecurity solutions. Austin Zhao, a senior research manager at IDC China, noted that while Mythos's capabilities were anticipated, its actual performance surprised many. Chinese companies are actively developing their own advanced models. 360 Security Technology, for instance, claimed to have developed an AI-powered "vulnerability discovery agent" that uncovered hundreds of previously unknown flaws, including those in widely used software such as Microsoft Office. Meanwhile, Chinese AI company Zhipu (also known as ZAI) launched its latest language model, GLM 5.2, just 24 hours after Anthropic's revelation about the US government's order. Research firm Artificial Analysis rated GLM 5.2 as the third most intelligent AI, trailing only Anthropic's Claude Fable and OpenAI's Chad GPT 5.5, but outperforming Google's Gemini. Notably, GLM 5.2 is significantly more cost-effective, priced at about 50 cents US per task compared to Claude Fable's nearly $3. Beijing-based consultancy Concordia AI indicates that while China's primarily open-source models currently lag behind US closed-source models in cyber capabilities, their progress has been rapid over the past year.

Emerging Global AI Dynamics

The rapid advancements in AI cybersecurity, exemplified by Mythos, are expected to drive up cybersecurity costs as companies invest more in personnel, infrastructure, and advanced protection systems. Conversely, this also creates new opportunities for AI-based security services. The G7 summit recently saw Western leaders engaging with the heads of prominent AI firms like Anthropic and OpenAI, highlighting the global focus on AI governance and development. AI researcher Jake Hersh Allen suggests that models like Zhipu's GLM 5.2 demonstrate that alternative methods exist for creating near-frontier AI, which do not necessarily require the same degree of financial investment or concentration of power seen in US-produced frontier models. This perspective challenges the notion of a "winner-take-all" industry, implying that international cooperation could foster more ethical or influential AIs. However, the competitive landscape is not without friction; Anthropic has accused Alibaba of illicitly extracting capabilities from its Claude AI models.

The global race for AI supremacy, particularly in critical areas like cybersecurity, is clearly intensifying, demanding careful navigation of both technological advancement and geopolitical realities.