Anthropic rejects Chinese think tanks’ access to its most powerful AI model, Mythos, further escalating the US-China AI game

null

Author: Claude, Deep Tide TechFlow

Deep Tide Guide: According to The New York Times on May 12, a Chinese think tank representative last month at a closed-door meeting organized by the Carnegie Endowment for International Peace in Singapore requested access to the Claude Mythos model from Anthropic, which was immediately denied.

The incident was subsequently reported to the White House, triggering high alert from the U.S. National Security Council.

Mythos is Anthropic’s most powerful AI model released in April this year, regarded as “digital weapon-grade” technology due to its offensive and defensive capabilities in cybersecurity. Currently, it is only accessible to about 40 U.S. and U.K. institutions. At the time of the incident, the Trump administration was drafting an AI regulatory executive order, and this week he will also visit China with a business delegation to discuss AI-related issues.

A closed-door dialogue in Singapore is becoming the latest flashpoint in the U.S.-China AI competition.

According to The New York Times on May 12, during an unpublicized meeting organized by the Carnegie Endowment for International Peace in Singapore last month, a Chinese think tank representative requested during a break: hoping the company would relax policies to allow Chinese parties access to its latest, most powerful AI model, Claude Mythos.

Anthropic immediately refused.

This was not an official diplomatic request from the Chinese government. However, according to multiple media reports, after the event was reported back to Washington, officials from the Trump administration’s National Security Council (NSC) became highly alert, viewing it as another signal of China’s ongoing pressure in the AI field.

Mythos: A “Digital Weapon” Far Superior to Its Predecessors, Restricted from Release

To understand the significance of this event, one must return to Mythos itself.

Claude Mythos Preview was officially released on April 7, 2026, but it was not open to the public. Anthropic limited its access within a cybersecurity defense initiative called “Project Glasswing,” only opening it to about 40 institutions, including Amazon, Apple, Microsoft, CrowdStrike, Cisco, Nvidia, JPMorgan Chase, and the Linux Foundation.

According to Anthropic’s official blog and TechCrunch reports on April 7, Mythos discovered thousands of zero-day vulnerabilities during internal testing—security flaws previously unknown to developers—covering all mainstream operating systems and browsers, with some vulnerabilities existing for as long as 27 years. In cybersecurity assessments like CyberGym, Mythos significantly outperformed its predecessor, Claude Opus 4.6. Its SWE-bench verification score reached 93.9%, compared to 80.8% for Opus 4.6.

China Excluded, Labeled as an “Adversarial Nation”

Anthropic classifies China as an “adversarial nation,” and its services are already unavailable within mainland China. The restricted release of Mythos explicitly excludes Chinese institutions.

According to a three-part series of reports from South China Morning Post from late April to early May, China’s response to the Mythos incident has been complex. Officially, China has been relatively restrained, with no major public statements or fierce responses. Some Chinese AI insiders even questioned whether Anthropic was creating a marketing stunt under the guise of security risks, limiting model access to U.S. companies.

However, reactions within the cybersecurity industry have been quite different. After Mythos was released, stocks of Chinese cybersecurity listed companies such as Qihoo 360, Sangfor, and 360 Security saw continuous gains for several days, with market expectations that AI-driven cybersecurity demand will accelerate.

Austin Zhao, senior research manager at IDC China, told South China Morning Post that China will “certainly see” Mythos-level models domestically, but the overall capability of current Chinese cybersecurity models “still lags far behind Mythos.” Nonetheless, Chinese models are rapidly improving, and this trend is irreversible. IDC predicts that the Chinese AI cybersecurity industry will grow from 1.58 billion RMB in 2025 to 59.35 billion RMB (about 8.7 billion USD) by 2030, an increase of over 37 times.

The real dilemma is that many Chinese banks, energy companies, and government agencies operate software systems that highly overlap with those vulnerable to Mythos discovering security flaws. But currently, China has no seat at this defensive upgrade table.

White House Alerts and Policy Battles: Drafting an Executive Order, Trump Visiting China This Week

The alert triggered by the Singapore closed-door meeting, combined with a series of larger policy battles.

According to The Washington Post on May 11, there are sharp disagreements within the Trump administration over AI regulation. On one side, national security officials—including NSA and the Office of the Director of National Intelligence—advocate for security assessments by intelligence agencies before AI models are publicly released; on the other side, the Department of Commerce prefers to keep assessment authority within its own jurisdiction. Kevin Hassett, director of the White House National Economic Council, revealed in an interview with Fox Business last week that the government is studying the issuance of an executive order to provide a clear roadmap for AI model security assessments, similar to the FDA’s pre-market review process for drugs.

Meanwhile, Trump is scheduled to visit China this week, where AI-related topics are expected to be discussed.

According to Axios on May 12, U.S. officials hope to “use the leader’s meeting to open a dialogue and see whether an AI affairs communication channel should be established.” However, Melanie Hart, senior director of the Atlantic Council’s China Center, warned that during the Biden administration, China mainly “collected U.S. information rather than seriously discussing AI protection” in AI security dialogues, and the officials attending the talks often lacked AI technical expertise.