NEW DELHI: US restrictions on access to advanced artificial intelligence (AI) models are beginning to reshape India’s AI strategy, with the government building independent capability to test critical software and cybersecurity systems instead of waiting for access to Anthropic’s most advanced model.
The ministry of electronics and information technology (MeitY) has started using alternative frontier AI models to stress-test critical code while accelerating work on indigenous frontier models, reflecting a broader effort to reduce reliance on overseas AI platforms for strategic applications. The move comes even as India continues to seek access to Anthropic’s Mythos 5 under Project Glasswing, a trusted-partner programme that Washington is yet to expand.
“We are one of the trusted countries to explore participation in Anthropic’s Glasswing exercise so that we have access to these models and can stress-test our own systems. At the same time, we cannot wait for that process. We have developed in-house capabilities to test critical code using other AI models that are currently around 60-70% as capable as Mythos,” MeitY Secretary S Krishnan said at the CII Cybersecurity Summit recently. He said the ministry is encouraging government departments to earmark at least 15% of their IT budgets for cybersecurity, signalling that strengthening cyber resilience is becoming a parallel priority as AI adoption gathers pace.
The government’s response follows Washington’s decision to retain export restrictions on Mythos 5 even after lifting curbs on Anthropic’s Fable 5 model for India. While access to Fable 5 resumed on July 1, Mythos 5 is available only to the original Project Glasswing participants.
Officials say the episode underscores the risks of relying on overseas firms for technologies that are central to national cybersecurity. The emphasis, therefore, is shifting from securing access to a single foreign model to ensuring India has the capability to evaluate, deploy and secure AI systems critical to government infrastructure.
The uncertainty has also strengthened the case for accelerating indigenous frontier AI development under the IndiaAI Mission, alongside efforts to deepen partnerships with trusted technology providers abroad. Officials maintain that engagement with Anthropic and the US government will continue, but say domestic capability can no longer remain contingent on external approvals.
Industry experts say India should also seek a more predictable framework for accessing strategic AI technologies.
“India should not settle for discretionary access. We should ask for allied-tier treatment, or at least a transparent, trusted-partner framework based on Pax Silica and similar agreements,” Nikhil Narendran, partner at Trilegal, told FE.
Jaspreet Bindra, co-founder of AI&Beyond, said the present arrangement risks making access to frontier AI a geopolitical bargaining chip. “Granting access on a case-by-case basis creates uncertainty and weakens the notion of a consistent ‘ally’ framework. India should advocate for transparent, tier-based eligibility similar to Nato+ or Quad-style technology partnerships rather than discretionary approvals. Otherwise, AI access risks becoming a bilateral bargaining tool,” he said.
The uncertainty is already influencing enterprise AI adoption. Companies are increasingly evaluating open-weight models such as GLM-5.2 and Qwen, which can run on domestic infrastructure at lower cost while offering greater flexibility for fine-tuning and deployment.
“Models such as GLM-5.2 and Qwen can run significantly cheaper on domestic infrastructure, eliminate recurring API costs, and allow complete fine-tuning. However, the US may still interpret widespread adoption as a geopolitical signal,” Bindra said.
Industry estimates place India as Anthropic’s second-largest market by consumption. But experts say the latest restrictions may have effects that outlast the export controls themselves. Enterprises that migrate to alternative models are unlikely to switch back quickly even if access to Mythos 5 is restored, as doing so would require fresh performance testing, retraining and regulatory certification.
Source: The Financial Express
