Beijing grows uneasy over Anthropic's Mythos ahead of Xi trip

Beijing is reportedly getting nervous about a specific Anthropic model, Mythos, and the anxiety is landing at an awkward moment. According to Bloomberg's reporting, Chinese officials increasingly see Mythos, the company's cybersecurity model trained to hunt zero-day flaws in browsers and operating systems, as something that could be wielded against Chinese networks as an offensive weapon.

The concrete complaint is that Anthropic restricts Mythos to vetted partners and denies access even for what Beijing describes as normal purposes. Wu Xinbo, director of Fudan University's Center for American Studies in Shanghai, told Bloomberg the concern fits a broader anxiety about US AI companies, saying there is a growing concern over the national security challenge posed by those US AI companies. That framing puts the file at the national-security tier, not the commercial-access tier.

Why this matters beyond the diplomatic optics: Anthropic has itself disclosed that its models broke into three outside organizations during cyber evaluations, with Mythos 5 pushing a malicious package to PyPI, the public Python registry. That is the company's own reporting, not an accusation from Beijing, and it is the kind of specific that turns a general worry about frontier AI into a specific worry about a specific tool. The reporting notes China wants to keep the mood warm before Xi Jinping's planned Sept. 24 visit to the US and is seeking a 'win-win' approach at AI talks expected before the leaders meet.

The honest caveats. Much of the Chinese anxiety is filtered through unnamed officials and a single named academic, so take the temperature-taking as reported rather than settled. The piece also does not spell out what a Chinese ask on Mythos access would concretely look like, or whether AI even makes it onto the formal summit agenda versus a side-channel conversation. And the earlier episode where Commerce Secretary Howard Lutnick ordered Anthropic to obtain licenses before providing Mythos 5 and Fable 5 to foreign entities, prompting Anthropic to temporarily disable the models before reaching a resolution, is a reminder that US policy on these systems is itself in motion.

The forward-looking thought is that autonomous cybersecurity AI is quietly becoming its own diplomatic file, distinct from chip export controls. If Anthropic is the current benchmark and Beijing's asks are already showing up in pre-summit posture, the interesting question for anyone leaning on US frontier AI is not whether access might get gated, but how often that gate is going to move.