Anthropic rejects Chinese government's request for access to its latest AI model, 'Claude Mythos Preview'.
Anthropic's '
China Thought Access to Anthropic's Newest AI The Answer Was No. - The New York Times
https://www.nytimes.com/2026/05/12/us/politics/china-ai-anthropic-openai-mythos-chatgpt.html
Claude Mythos Preview has an extremely high ability to discover and exploit software vulnerabilities, reaching a level that, according to Anthropic, 'outperforms most humans, with the exception of a handful.' It has also achieved a high success rate of 72.4% in creating exploits that take advantage of vulnerabilities, and has launched 'Project Glasswing,' an urgent initiative to use AI for defensive purposes before attackers begin to exploit vulnerabilities, by granting access to AI models to specific organizations to help discover them.
Anthropic develops 'Claude Mythos Preview,' an AI with extremely high cyberattack capabilities, and has also launched 'Project Glasswing,' which will provide a preview version to Microsoft, Apple, and others - GIGAZINE
Anthropic has recognized the risks of Claude Mythos Preview and has refrained from making it publicly available, instead limiting its release to specific organizations such as some software companies and government agencies. However, it has also been reported that a small number of users have illegally obtained access to Claude Mythos Preview.
Reports indicate that access to 'Claude Mythos Preview,' which is only available to a select group of organizations, has fallen into the hands of unauthorized users - GIGAZINE
It has come to light, according to sources, that a Chinese think tank has requested access to the Claude Mythos Preview. A source who spoke to The New York Times said that at a meeting held in Singapore in April 2026, a representative of the Chinese think tank approached executives at Anthropic and strongly urged them to grant Chinese authorities access to the Claude Mythos Preview.
While this request was not an official demand from the Chinese government, sources say the meeting in Singapore was part of an exchange aimed at promoting dialogue between the United States and China.
According to sources, Anthropic has refused the request.
Matt Sheehan, a senior fellow at the Carnegie Endowment for International Peace and the organizer of the meeting held in Singapore, declined to comment on the content of the discussions at the conference, but said, 'This conference is essential for both the United States and China. It is extremely important for experts from each country to communicate about the potential risks that AI poses, and that is why we provide this forum for dialogue.' The New York Times also reached out to representatives of Anthropic for comment, but did not receive a response.
A U.S. government official commented, 'While members of a Chinese think tank may have lobbied, it is almost certain that the Chinese government approved and directed the message.' It is believed that Chinese authorities have strong control over think tanks, and are even more likely to intervene when it comes to informal diplomacy like this.
Furthermore, several companies in China are developing high-performance AI models, and in the field of open models, where models are made publicly available, they are showing results that surpass those of American companies. The Chinese-made open model 'DeepSeek V4 Pro,' which appeared in April 2026, has been shown to have performance equivalent to GPT-5 in performance analysis by a US government agency, and has been evaluated as 'China's cutting-edge model being eight months behind America's cutting-edge model.'
'DeepSeek V4 Pro is about eight months behind major US AI models, but is currently the most high-performing Chinese-made AI model,' reports CAISI, the US government's AI risk management agency - GIGAZINE
On the other hand, according to The New York Times, some U.S. government officials and industry insiders believe that OpenAI's ChatGPT 5.5 and Anthropic's Claude Mythos Preview could have significantly extended the U.S. lead over China in AI development to between nine months and a year. Furthermore, Chinese media outlets have pointed out that 'Claude Mythos Preview poses a significant risk to Chinese companies, and Anthropic's limited scope of operations is creating a technical gap,' while the Chinese side has argued that 'it is a mistake for Anthropic and OpenAI to strictly control the models, and China needs access to these models in particular to protect its critical infrastructure.' These reports suggest that the Chinese government is highly wary of falling behind in the intensifying AI development race.
Related Posts:
in AI, Posted by log1e_dh