Last month in Singapore, during a closed-door meeting organized by the Carnegie Endowment for International Peace, a Chinese think tank representative made a request to Anthropic: to open access to their most powerful model, Claude Mythos.

Anthropic refused on the spot.
After the event was reported back to Washington, the White House National Security Council became highly alert, viewing it as a signal of China’s continued pressure in the AI field.
Mythos is Anthropic’s most powerful model to date, but it is not open to the public. It autonomously discovered thousands of zero-day vulnerabilities during internal testing, covering all mainstream operating systems and browsers, with some vulnerabilities existing for as long as 27 years. The SWE-bench validation score is 93.9%, compared to 80.8% for the previous Opus 4.6.
Anthropic limited its access within the "Project Glasswing" cybersecurity defense framework, opening it only to about 40 U.S. and U.K. institutions, including partners such as Amazon, Apple, Microsoft, Nvidia, and JPMorgan Chase.
Anthropic explicitly designated China as an "adversarial nation," excluding Chinese institutions from the limited release of Mythos.
The real dilemma is that the underlying software running many Chinese banks, energy companies, and government agencies overlaps heavily with the systems where Mythos discovered vulnerabilities. The vulnerabilities have been found, but China cannot obtain this defensive advantage.
After Mythos was released, Chinese cybersecurity listed companies like Qihoo 360, Sangfor, and 360 Security saw their stock prices rise for several consecutive days, with the market reacting faster than any statement. IDC predicts that the scale of China’s AI cybersecurity industry will grow from 1.58 billion yuan in 2025 to 59.35 billion yuan in 2030, an increase of over 37 times.
Meanwhile, within the Trump administration, a struggle is underway over AI regulation. The national security system advocates for security assessments by intelligence agencies before AI model releases, while the Ministry of Commerce wants to retain the assessment authority.
Trump is visiting China this week, and AI topics are expected to be part of the discussion. But Melanie Hart, senior director at the Atlantic Council, warned: during the previous Biden era, China’s main goal in AI security dialogues was "collecting U.S. information rather than seriously discussing AI protection."
A request at a closed-door meeting being refused has triggered a chain reaction: the weaponization potential of the most powerful AI models, the deepening of U.S.-China technological decoupling, and the reality that China’s critical infrastructure is being excluded from AI cybersecurity upgrades.