India considers making top AI developers test safety before launch

A tougher track for the biggest models

India is looking at a separate safety framework aimed only at so-called frontier models. These are the newest and most capable AI systems from the leading labs. Under the idea, developers might have to show that proper safeguards exist before they release a model that could cause serious harm.

As Crypto Briefing reported, the story first came out through NDTV, which spoke to officials familiar with the talks. The outlet describes it as a possible split into two lanes, with ordinary AI tools handled more lightly and the most powerful systems facing extra scrutiny.

That would be a shift. So far India has leaned toward encouragement over obligation, and a mandatory pre-launch step would be a much firmer approach.

What the checks could look like

The report lists several possible layers. None of them is settled, and they are described as ideas under review rather than rules.

• Pre-release evaluations: examining a model's abilities before the public can use it.
• Controlled-environment testing: letting dangerous capabilities show up in a lab setting instead of out in the real world.
• Sandbox requirements: a sandbox is a supervised space where developers can try a model out under watch.

The worries behind this are specific. Frontier models are seen as able to help carry out cyberattacks and to produce very realistic fake content. The report does not say who would run the tests, what pass marks would be, or which companies would be covered.

The consultation paper and the timeline

The next step is a consultation paper, a document that sets out proposals and invites reactions before a law is written. Union Electronics and IT Minister Ashwini Vaishnaw has said it will appear in the coming weeks. It is meant to gather views from industry, civil society, citizens and other stakeholders.

The source also points to a possible arrival as early as November 2026, and places the minister's comments between October 23 and 28, 2026. Treat those dates as approximate, since the government has only said "coming weeks."

The paper is expected to shape India's forthcoming AI law, so what it says about frontier models will set the terms of the debate.

From voluntary guidance to possible law

In November 2025, India published its AI Governance Guidelines. That document favored voluntary, risk-based steps. It admitted that AI risks were growing, but it stopped short of calling for a standalone AI law. Within a year, talk has turned to formal legislation with special duties for the most capable systems.

Some groundwork exists. The IndiaAI Safety Institute is already running, and its oversight role may grow as discussions go on. Security officials have also raised concerns. CERT-In, the country's computer emergency response team, has issued warnings about agentic AI, meaning systems built to take actions for a user instead of just answering questions.

The source adds that Anthropic has reportedly talked with the Ministry of Electronics and Information Technology (MeitY) about model access and safety testing. The company is said to favor mandatory transparency rules. The source describes this only as reported, not confirmed.

Who feels it, and what to watch

The first impact would fall on AI developers. If evaluations and sandbox testing became compulsory, companies would need to set aside time and money for them. That could raise costs and slow product launches, though the size of the burden depends on the fine print.

The most important fine print is the definition of a frontier model. A narrow definition would catch only a handful of labs; a broad one could reach many more. The source also says the approach fits with international trends and draws on US-style thinking about advanced AI.

For ordinary users, the effect would be indirect. New models might reach India later, or with limits, if review becomes a required step. Three things will show where this goes: the consultation paper and its definition, any expansion of the Safety Institute's powers, and how industry responds.