India weighs pre-release safety checks for frontier AI models

India may soon ask the builders of its most powerful AI systems to prove those systems are safe before they ship.

The government is examining a separate safety framework for so-called frontier models. Developers could have to demonstrate adequate safeguards before releasing systems capable of serious harm. That would mark a real change for a country that has so far preferred gentle nudges to hard rules.

The plan surfaced in a report by NDTV, which cited officials familiar with the discussions. It is tied to a consultation paper on AI regulation that the government is preparing. That paper will shape India’s forthcoming AI law.

Two lanes for AI

The framework under consideration may include several layers of scrutiny, according to the research findings:

• Pre-release evaluations to assess what a model can do before it reaches the public.
• Controlled-environment testing, so risky capabilities surface in a lab rather than in the wild.
• Sandbox requirements, which would let developers trial models under supervision.

None of this is final. The measures are being examined, not enacted, and the consultation process exists precisely to test them against industry and public feedback.

The risks driving the conversation are concrete. Frontier models are seen as capable of enabling cyberattacks and generating highly realistic synthetic content.

A consultation paper in the coming weeks

Union Electronics and IT Minister Ashwini Vaishnaw has said the consultation paper will be released in the coming weeks. It will seek input from industry, civil society, citizens and other stakeholders.

The research findings indicate the paper could arrive as early as November 2026. They place the minister’s announcement between October 23 and 28, 2026.

At least one major AI developer appears to have engaged already. Anthropic has reportedly held discussions with the Ministry of Electronics and Information Technology (MeitY) on model access and safety testing. The company has reportedly advocated for mandatory transparency rules.

From voluntary guidelines to legislation

The proposal represents a notable pivot in India’s approach.

In November 2025, the government issued the India AI Governance Guidelines. That document emphasized voluntary and risk-based measures. It acknowledged that AI risks were escalating but stopped short of recommending a standalone AI law.

Less than a year later, the conversation has moved to formal legislation with distinct obligations for the highest-capability systems.

Some institutional groundwork is already in place. The IndiaAI Safety Institute is operational. According to the research findings, its oversight role may be expanded as the regulatory discussions progress.

Security agencies have also been sounding alarms. Authorities including CERT-In, India’s computer emergency response team, have issued warnings and advisories about risks linked to agentic AI. Those are systems designed to take actions on a user’s behalf rather than simply answer questions.

What this means for AI developers and investors

For companies building frontier systems, the most immediate implication is compliance. If pre-release evaluations and sandbox testing become mandatory, developers would need to budget time and resources for them. That could add operational costs and stretch product deployment timelines.

The size of that burden will depend on the details. How a frontier model gets defined will be one of the most consequential lines in the eventual law.

There is also a competitive dimension. The research findings note that India’s approach aligns with international trends in AI regulation and draws inspiration from US-style approaches to advanced AI.

The key items to watch are now fairly clear. The first is the release of the consultation paper and its precise definition of a frontier model. Next comes the scope of any expanded mandate for the IndiaAI Safety Institute. Finally, there is the response from industry, which will signal how much of the proposed framework survives contact with the companies it targets.