India’s access to Anthropic’s frontier AI model Mythos has been delayed as the government and the US-based company remain divided over how the model should be tested, with New Delhi insisting that evaluation be carried out on infrastructure located within the country, according to official sources.
Anthropic had initially proposed that Indian entities route their code through its cloud infrastructure for testing, an arrangement the government was unwilling to accept because of concerns around data residency and the handling of sensitive information. Instead, India has sought a deployment model that allows testing to take place on infrastructure hosted within the country.
“Anthropic had proposed that you send the code across to a cloud setup and do your testing there. We were a little reluctant to do that. We want to do it in India,” an official said.
The disagreement reflects India’s broader approach towards sovereign AI infrastructure, where sensitive government or enterprise data is expected to remain within the country’s jurisdiction.
Officials said the preference is to deploy or evaluate frontier AI models on infrastructure located in India rather than rely on overseas cloud environments where data could potentially move outside the country’s control.
However, implementing such an arrangement is not straightforward. Since Mythos is a closed proprietary model, Anthropic is unlikely to release the model weights that would allow customers to run it independently on their own servers. The alternative would require the company to deploy the model on infrastructure physically located in India while retaining operational control, a deployment model that has yet to be finalised.
The delay comes even as earlier regulatory hurdles in the US have eased. In June, the US Department of Commerce had imposed export controls that require Anthropic to obtain government approval before granting foreign nationals access to Mythos and another frontier model, Fable. The restrictions were introduced after researchers reportedly identified a way to bypass Fable’s safety guardrails.
Those export controls were lifted on June 30, allowing Anthropic to restore access to its existing Project Glasswing partners in the United States. While Fable has since become available more broadly, officials said Mythos has not yet been rolled out internationally.
India has already submitted a formal request identifying the government departments and companies it wants cleared for access under Project Glasswing, Anthropic’s security initiative that uses Mythos to identify vulnerabilities in critical systems. Officials said the company has not indicated a timeline for extending Mythos access to India, with discussions on the deployment framework still continuing.
The delay means India’s participation in the programme now hinges not only on regulatory clearances but also on reaching an agreement over where and how the model will be deployed, officials said.