The geopolitical battle over artificial intelligence has entered a new phase, moving beyond hardware export controls to the direct restriction of access to frontier software models. The latest flashpoint centers on Anthropic’s highly anticipated, yet unreleased, “Mythos” model.
According to a report by The New York Times, summarized by The Times of India, a representative from a Chinese think tank formally requested early access to Mythos at a closed-door meeting in Singapore in April 2026. The meeting was hosted by the Carnegie Endowment for International Peace as part of a Track 2 diplomatic dialogue. Anthropic refused the request.
The exchange triggered significant alarm within the White House National Security Council. U.S. officials told the Times that the request had “almost certainly been approved by Beijing,” indicating that the inquiry was not an individual initiative but a sanctioned effort by the Chinese government to evaluate the model’s capabilities.
What Makes Mythos Different
Anthropic unveiled Mythos on April 7, 2026, but has not released it to the public. Access has been limited to the U.S. government and approximately 40 partner organizations. The reason for the restricted release is the model’s extraordinary capabilities in the cybersecurity domain.
According to the Times, Mythos can find previously unknown vulnerabilities in every major operating system and web browser. More alarmingly, it can write working exploits for those vulnerabilities without human guidance. This combination of vulnerability discovery and autonomous exploit generation represents a qualitative leap in the potential for AI-enabled cyberattacks.
As we reported last month, the emergence of Mythos triggered emergency diplomatic discussions about the implications of such capabilities for global cybersecurity. The model’s potential to automate attacks on critical infrastructure, power grids, water systems, financial networks, has alarmed officials across multiple agencies.
The White House Response
The alarm within the NSC was not limited to the think tank request itself. In April, Vice President JD Vance personally called the CEOs of leading AI companies, including Sam Altman of OpenAI and Dario Amodei of Anthropic, to express concern about the potential for Mythos-class models to be used against small-town banks, hospitals, and water treatment plants.
Treasury Secretary Scott Bessent separately warned top banking executives about the risks posed by frontier AI models with autonomous cyberattack capabilities. The White House is now reportedly weighing an executive order that would establish a formal review process before the most powerful AI models can be shared with any foreign entities.
The Pentagon, which has publicly declared Anthropic a supply-chain risk and is in the process of reducing its reliance on the company, has nonetheless been using Mythos to patch vulnerabilities in federal systems, a striking illustration of the dual-use dilemma at the heart of frontier AI development.
The Widening Gap and Its Implications
The episode has also reignited debate about the size of the AI capability gap between the United States and China. U.S. officials had previously estimated the American lead at approximately six months. With the emergence of Mythos and other frontier models, the Times reported that some officials now believe the gap may have widened to closer to a year.
This assessment is contested. Chinese cybersecurity stocks, including Qi An Xin and 360 Security Technology, rallied sharply after news of Mythos’s capabilities became public, suggesting that Chinese investors view the model as a genuine threat to domestic cyber defenses. A Chinese analyst quoted by the Times described the asymmetry as “swords on one side, a Gatling gun on the other.”
The Chinese state broadcaster CCTV’s Yuyuan Tantian account described Mythos’s capabilities as “a genuine threat,” a rare acknowledgment from Chinese state media of a significant U.S. technological advantage.
The request for access to Mythos, and the White House’s alarmed response, illustrates the fundamental tension at the heart of U.S. AI policy. Washington wants to maintain its technological lead while also engaging with Beijing on AI safety and risk mitigation. But if the most capable AI models are also the most dangerous, the question of who gets access to them will remain one of the defining geopolitical questions of the coming decade.
The Mythos episode also raises a deeper question about the nature of AI safety as a diplomatic concept. The U.S. government has been pushing for international AI safety frameworks, including commitments to human oversight and the prevention of catastrophic misuse. But those frameworks are difficult to negotiate credibly when one side possesses a model that the other side is actively trying to access and evaluate. The Singapore meeting, and the alarm it triggered in Washington, suggests that the gap between AI safety as a diplomatic aspiration and AI capability as a geopolitical weapon is narrowing, and that the two may be impossible to fully separate.
(Related: Mythos Triggers US-China AI Emergency Channel Talks Ahead of Trump-Xi Summit)