New Delhi: Anthropic is in talks with the Ministry of Electronics and Information Technology (MeitY) to make its AI models available to the IndiaAI Safety Institute for safety testing, the company’s India policy head Amlan Mohanty said Thursday.
“We have been having discussions about how we would make Anthropic’s models available to them for safety testing. The intent is entirely there,” Mohanty said at the closing plenary of the Trust and Safety India Festival (TASI) 2026 in New Delhi.
He said Anthropic has also asked MeitY for binding rules that require AI companies to be transparent about their models. “We need binding regulations for AI companies, whether they’re foreign or Indian, to be very transparent about their models,” he said. He said he had made the case to a senior MeitY official earlier in the day.
The IndiaAI Safety Institute was announced in January 2025 under the IndiaAI Mission. It runs on a hub-and-spoke model with academic institutions, startups and industry, and tests AI systems for risks in India’s linguistic and social context.
In July, MeitY said it was in talks with US counterparts to access Anthropic’s model for cybersecurity testing of India’s critical digital infrastructure. That month, MeitY Secretary S. Krishnan said export restrictions on Anthropic’s Mythos model had been eased, and that Anthropic wanted to extend access to entities in 15 countries, including India.
India is Claude’s second-largest market. Anthropic opened its Bengaluru office in February.
The call for binding rules comes as the Centre prepares a dedicated AI law. In November last year, MeitY’s India AI Governance Guidelines ruled out an immediate need for a standalone AI law. In July, Krishnan said “probably the time has come now to look at a separate legislation for AI”, and that the ministry would prepare a draft.
At present, India’s only binding AI-specific rules are the February amendments to the IT intermediary rules, which mandate labels on AI-generated content.
‘Reject the notion’
Mohanty’s remarks come two days after a Reuters report that said Anthropic and OpenAI had sought significant control over what outside evaluators embedded in their labs could publish, citing trade secrets. In September, Anthropic CEO Dario Amodei had said evaluators should be able to publish key findings without editorial control. Anthropic told Reuters the report did not accurately characterise all its discussions.
“Evaluators have badge access. They can sit anywhere in the office, access proprietary information and talk to employees,” Mohanty said. “I reject the notion that we are not giving them the ability to do that.”
Anthropic named Accenture’s Faculty unit its first embedded evaluator on 18 September, with each company committing at least $1 billion over five years.
Mohanty said the line should be drawn at model weights. “That is our trade secret. I don’t understand why anybody would need access to it,” he said.
‘Nothing like content’
Mohanty said safety systems built for content moderation cannot deal with AI agents that work over long periods, interact with other agents and act in the real world. “This is nothing like content. We need a completely new paradigm,” he said.
“Being inside a frontier AI lab changed my view. You have to think in exponentials,” said Mohanty, who joined Anthropic in April. “We don’t have the answers to these questions.”
He said cyber defenders must get the most capable models first, leaving attackers “one generation behind”. He added, “The default trajectory is that most of the world will not benefit from AI unless we are very intentional about it.”
1.6 years to 1.6 days
Rajesh Ranjan, Head of Government Affairs and Public Policy at Google, said the time taken to discover a software vulnerability has fallen from 1.6 years to 1.6 days. “The defender has to be right every single time. The attacker needs to be right once,” he said.
Google uses one AI agent to find vulnerabilities and another to patch them. Its Big Sleep agent detects flaws, and CodeMender, launched last October, writes the fixes.
Ranjan said Google’s SynthID watermark can be detected even if a user photographs an AI-generated image. On 7 October, Google opened its SynthID Detector to the public. It reads content from OpenAI, Nvidia and Kakao, with Apple to follow.
People who cannot read English are most at risk, Ranjan said. “Anyone who cannot understand the content well is more vulnerable,” he said.
TASI was hosted by the Centre for Social Research and the Trust & Safety Forum at the India International Centre.
(Edited by Viny Mishra)
Also read: AI models could soon replicate themselves across smart TVs, fridges, Anthropic warns