The following original contribution for Friends of Socialist China is by İbrahim Can Eraslan, a Turkish socialist and student of Chinese Law and Governance at Tongji University in Shanghai. He examines a striking development in the global technology race: the emergence of Chinese artificial intelligence systems that now rival the most advanced Western models in the strategically sensitive field of cybersecurity – achieved, remarkably, under conditions of escalating US technological warfare.
Eraslan shows how Washington’s campaign of export controls and suppression is producing the opposite of its intended effect. By restricting access to the United States’ AI systems – even shutting down Anthropic’s frontier models worldwide at three days’ notice – the US is driving governments and enterprises across the Global South towards cheaper, open-weight Chinese alternatives that cannot be embargoed or switched off by executive order. In closing the gap through algorithmic ingenuity and open development rather than raw computing power, he argues, China’s socialist system is once again outcompeting monopoly capitalism in the technologies that will define the twenty-first century.
It is, Eraslan concludes, a further step towards a multipolar technological order in which no single imperialist bloc can monopolise the most consequential technologies of our era.
In a development that has sent shockwaves through Washington’s national security establishment, Chinese artificial intelligence laboratories have achieved near-parity with the United States’ most advanced cybersecurity AI systems — and they have done so under conditions of escalating American technological warfare. Zhipu AI’s open-weight GLM-5.2 model and 360 Security Technology’s Tulongfeng system have now demonstrated capabilities that match Anthropic’s tightly controlled Mythos in the automated discovery of software vulnerabilities, fundamentally altering the strategic calculus of the global AI race.
The implications extend far beyond the technical domain. These achievements come precisely as the Trump administration has intensified its campaign of technological suppression against Chinese firms — a campaign that is now producing the exact opposite of its intended effect. Instead of maintaining American dominance, Washington’s export restrictions are accelerating China’s indigenous innovation while simultaneously alienating global users who are increasingly turning to Chinese open-weight alternatives.
From Zero to Parity Under Siege
Zhipu AI’s GLM-5.2, released between June 13 and 16, 2026, represents a genuine milestone in open-source artificial intelligence. With 753 billion parameters and a one-million-token context window, the model operates under a fully permissive MIT license — meaning any individual, organisation, or government can download, modify, and deploy it without restriction, without geographic limitations, and without dependency on a commercial API controlled by a San Francisco corporation.
What makes this release strategically significant is not merely its technical specifications but its demonstrated performance in the highly specialised domain of cybersecurity. Security researchers evaluating GLM-5.2 against Anthropic’s Claude Opus 4.8 — the latter being the most capable publicly available Western model — found that the Chinese system matched or exceeded its American counterpart on key vulnerability-detection benchmarks. Semgrep’s independent testing placed GLM-5.2 ahead of Claude Opus 4.8 in certain scenarios, while the model has rapidly ascended to become one of the ten most-used AI systems globally, according to data from OpenRouter.
Simultaneously, 360 Security Technology (Qihoo 360) unveiled its Tulongfeng (“Dragon-Slaying Blade”) platform at the ISC.AI 2026 conference in Beijing on June 28. Founder Zhou Hongyi presented the system as China’s direct answer to Mythos — not as a single monolithic model, but as a multi-agent swarm architecture that orchestrates specialised AI agents across the entire vulnerability research pipeline, from threat landscape modelling and data-flow analysis to automated sandbox construction and exploit confirmation.
The numbers Zhou presented were striking: Tulongfeng has reportedly identified 3,432 software vulnerabilities since deployment, of which 105 have been officially confirmed by Chinese government authorities. Among the claimed discoveries are a Windows kernel privilege escalation vulnerability dormant for five years, an Office remote code execution flaw dormant for eight years, and an Excel vulnerability dormant for a decade — all allegedly earning recognition from Microsoft itself.
To be sure, these specific claims await independent verification, and some have already been disputed. Eugenio Benincasa, a senior researcher at ETH Zurich’s Center for Security Studies, concluded in April 2026 that Qihoo 360’s AI capabilities, while significant, “do not yet appear to match the reasoning capabilities described for Claude Mythos.” Yet even accounting for a margin of promotional exaggeration, the underlying trajectory is unmistakable: Chinese firms are systematically closing the gap in one of the most strategically consequential applications of frontier AI, and they are doing so through an approach that emphasises practical engineering over raw computational brawn.
Zhou himself acknowledged that Chinese frontier models still trail their American counterparts by 20 to 30 percent in base capability. Rather than attempting to close this gap through sheer compute — a race that US chip export controls are designed to make unwinnable — 360 Security Technology built around it. “If the American approach is about cultivating a genius hacker,” Zhou told the Beijing conference, “the 360 approach is about organising a professional attack-and-defence team.”
This is not merely a different technical strategy. It is a fundamentally different philosophical approach to technological development — one that prioritises systemic resilience, mass accessibility, and operational continuity over the concentration of capability in proprietary black boxes controlled by a handful of US corporations.
The Self-Defeating Logic of American Suppression
The timing of these Chinese breakthroughs could scarcely be more instructive. The very week that Zhipu AI and 360 Security Technology were demonstrating their cybersecurity capabilities, the Trump administration’s restrictions on American AI exports were producing a cascade of unintended consequences that even some former US officials are now warning against.
On June 12, 2026 — just three days after Anthropic launched its most advanced models, Claude Fable 5 and Claude Mythos 5 — the US government issued an export-control directive ordering Anthropic to suspend access to both systems for all foreign nationals worldwide. Because Anthropic possessed no technical mechanism to filter foreign nationals from domestic users in real time, the practical effect was an immediate global shutdown. The models, which had represented the pinnacle of American AI capability, went dark for every customer on Earth within hours.
The government’s stated rationale — that a purported “jailbreak” technique could bypass Fable 5’s cybersecurity safeguards — was publicly disputed by Anthropic itself, which characterised the concern as a “misunderstanding” based on vulnerabilities that “other publicly available models can also surface.” Yet the damage was done. For organisations around the world that had built operational pipelines around Anthropic’s models, the message was received with perfect clarity: American AI infrastructure is politically unreliable.
The suspension has since been partially lifted — Mythos 5 access was restored in a limited capacity after approximately two weeks — but the broader signal has not been lost on the global market. When the United States demonstrates its willingness to shut down access to frontier AI models on three days’ notice, governments and enterprises in the Global South have every rational incentive to seek alternatives that cannot be severed at the whim of a Washington bureaucrat.
This is where Zhipu AI’s open-weight strategy acquires its geopolitical significance. GLM-5.2 cannot be shut down by executive order. Its weights are distributed across thousands of servers worldwide. It can run in air-gapped environments with no external network connection. For nations that value technological sovereignty — a category that includes the vast majority of countries outside the US-led bloc — this architectural resilience is not a peripheral feature. It is the central consideration.
The Trump administration’s restrictions have extended beyond Anthropic. OpenAI’s GPT-5.6 family — its newest model line — has been placed in limited preview accessible to only a handful of partner organisations. The practical effect has been to constrict global access to the most capable American AI systems at precisely the moment when Chinese alternatives are becoming competitive.
Saif Khan, a former adviser on export controls during the Biden administration and researcher at Georgetown University’s Center for Security and Emerging Technology, has been among the most articulate critics of this approach. The curbs, Khan argues, are achieving the diametric opposite of their stated purpose: they are pushing global users toward cheaper, open-weight Chinese alternatives while simultaneously eroding the competitive position of American AI developers who depend on international markets for revenue and influence. When the US government restricts access to its own companies’ products, it is effectively subsidising the market share of their Chinese competitors.
The Pentagon’s Panicked Response
The American national security establishment has not been blind to these dynamics, though its responses have been characteristically contradictory. While one branch of the US government restricts AI exports on security grounds, another has scrambled to bolster domestic alternatives — most notably through its backing of Reflection AI, a startup founded in 2024 by former Google DeepMind researchers.
In May 2026, the Pentagon announced agreements with eight commercial AI firms — including Amazon Web Services, Google, Microsoft, OpenAI, SpaceX, NVIDIA, Oracle, and Reflection AI — to deploy their models on classified military networks at Impact Level 6 (Secret) and Impact Level 7 (Top Secret/SCI). The inclusion of Reflection AI, a company with no prior defence contracting experience and no publicly released frontier model, was the most strategically revealing element of the announcement.
Reflection AI’s central value proposition to the Pentagon is precisely what makes it a telling symptom of American strategic anxiety: the company plans to release an open-weight frontier model trained on “tens of trillions of tokens” — explicitly positioned as a domestic counterweight to Chinese open-source models like DeepSeek and GLM-5.2. The Pentagon’s enthusiasm for this approach reflects a belated recognition that the open-weight paradigm is becoming the global standard, and that the US risks ceding the sovereign-AI market entirely to Chinese providers if it does not cultivate its own open-source champions.
The most conspicuous absence from the Pentagon’s list was Anthropic — the very company whose Mythos model represents the gold standard for AI-assisted cybersecurity. Defence Secretary Pete Hegseth designated Anthropic a supply chain risk in March 2026 after the company refused to permit Pentagon use of Claude for “all lawful” purposes, citing concerns about mass domestic surveillance and lethal autonomous weapons. The Pentagon subsequently cancelled a $200 million contract with Anthropic — the first time this designation has been applied to a US-based company — and the company responded by suing the federal government.
The spectacle of the US military simultaneously suppressing Anthropic’s exports on security grounds, excluding Anthropic from classified defence work on contractual grounds, and scrambling to fund a startup to replicate the Chinese open-weight model that is benefiting from both policies would be farcical if it were not so consequential. It is, in miniature, a portrait of an imperial power losing its strategic coherence.
What This Tells Us About Socialist Development
The Chinese AI advances in cybersecurity are not occurring in isolation. They are the latest manifestation of a broader pattern that has become impossible to ignore: China’s socialist system is systematically outcompeting Western capitalism in the technologies that will define the twenty-first century.
When the US launched its “chip war” against China — a campaign of export restrictions, Entity List designations, and technological blockades that has now targeted Zhipu AI (January 2025), Qihoo 360 (May 2020), and dozens of other Chinese technology firms — the explicit objective was to arrest China’s technological development by denying it access to advanced semiconductors and frontier AI models. The actual result has been to catalyse a wave of indigenous innovation that is producing cheaper, more efficient, and more accessible alternatives.
Consider the record. DeepSeek’s R1 model, developed for approximately $6 million, outperformed models that American companies had spent hundreds of billions to build — and DeepSeek released it as open source. GLM-5.2 matches or exceeds Claude Opus 4.8 on key benchmarks at roughly one-sixth the cost. Tulongfeng achieves vulnerability-detection parity not by outspending Mythos on compute but by architecting a multi-agent system that leverages two decades of proprietary security expertise. In every case, the Chinese approach has been characterised by algorithmic ingenuity, systems thinking, and a commitment to open development that contrasts sharply with the proprietary, profit-maximising model of Silicon Valley.
This is not coincidental. It reflects structural features of China’s socialist market economy that its Western critics consistently fail to comprehend. China’s state-led development model enables the coordination of massive resources toward strategic priorities — semiconductor independence, AI capability, cybersecurity resilience — without the constraint of quarterly earnings expectations or shareholder value maximisation. The results speak for themselves: when the Pivot to Asia began in 2011, the US led in 60 of 64 critical technologies globally; by 2022, China had surpassed the US in 52 of those 64 technologies (albeit this is something that can be measured in several different ways, and the statistic is being used by the hawkish Australian Strategic Policy Institute (ASPI) to foment anti-China hysteria).
Keith Bennett captured this dynamic in a recent interview when he observed that China’s growth is now expressed “above all in qualitative terms, with advances in AI, robotics and other fields all contributing to what many people in the world, young people in particular, are describing as ‘a very Chinese moment’ in their lives.” The emphasis placed by China’s incoming 15th Five-Year Plan on “new high quality productive forces” — including robotics, electric vehicles, and AI — provides further evidence that what is good for China’s socialist development is also good for humanity’s collective technological advancement.
Toward a Multipolar Technological Order
The cybersecurity AI breakthroughs of June 2026 carry implications that extend well beyond the technical community. They represent a further step toward a multipolar technological order in which no single nation — and certainly no single bloc of imperialist powers — can monopolise access to the most consequential technologies of our era.
When Zhou Hongyi described Anthropic’s Mythos as a “cyber nuclear weapon” and declared that “this kind of powerful weapon that can change the landscape of cyber offence and defence cannot be held only by others,” he was articulating a principle of technological sovereignty that resonates across the Global South. The United States has spent decades using its technological superiority as an instrument of coercion — extracting concessions, enforcing compliance, and punishing disobedience through the selective denial of access. The emergence of Chinese open-weight alternatives that cannot be sanctioned, cannot be embargoed, and cannot be shut down by executive order represents a fundamental challenge to this architecture of technological imperialism.
The Five Eyes intelligence alliance issued a warning during the same week as Qihoo 360’s announcement, cautioning that adversaries could begin using AI to conduct sophisticated cyberattacks “within months rather than years.” What the alliance did not acknowledge — perhaps could not acknowledge — is that the principal driver of this acceleration is not Chinese aggression but American overreach. By weaponising AI access, by demonstrating that American technology cannot be trusted to remain available, and by forcing the global community to seek alternatives, Washington has done more to democratise advanced AI capabilities than any open-source manifesto ever could.
Professor Jie Tang, the Tsinghua University researcher who founded Zhipu AI, predicted this week that a Chinese AI model reaching full Mythos-level capability will arrive before the first quarter of 2027 — a timeline that, if accurate, represents a dramatic compression of what most observers estimated earlier this year. Given the trajectory of the past six months, there is little reason to doubt him.
The AI race is not over. But it is no longer a race that the United States can expect to win through restriction, suppression, and technological warfare. The future belongs to those who build — and China’s socialist system has demonstrated, once again, that it can build with an efficiency, a scale, and a commitment to shared benefit that monopoly capitalism cannot match.
