Anthropic Discusses AI Safety Testing With MeitY, Calls for Binding Transparency Rules

Highlights

• Anthropic is in discussions with MeitY about making its AI models available for safety testing.

• The proposed evaluations would involve the IndiaAI Safety Institute.

• Anthropic’s India policy head has called for binding transparency rules covering Indian and foreign AI companies.

• India is considering a dedicated AI law as policymakers reassess existing governance measures.

• The discussions highlight questions around independent evaluation, model security and access to information needed to assess AI risks.

Anthropic Discusses Model Access for India’s AI Safety Institute

Anthropic is discussing arrangements with India’s Ministry of Electronics and Information Technology (MeitY) to make its artificial intelligence models available for safety testing, according to comments by Amlan Mohanty, the company’s India policy head.

Speaking at the closing plenary of the Trust and Safety India Festival 2026 in New Delhi on October 15, Mohanty said discussions were under way on providing model access to the IndiaAI Safety Institute. His remarks indicate an intention to support the process, although the announcement did not confirm a final agreement, a testing schedule or which models would be included.

The proposed engagement comes as India examines how to evaluate increasingly capable AI systems and develop governance arrangements suited to their potential risks. The IndiaAI Safety Institute, established under the IndiaAI Mission, is intended to support AI safety research, evaluation and testing.

Access to models can help evaluators examine how systems behave under different conditions, including their responses to potentially harmful requests and their performance in security-sensitive scenarios. The precise scope of any Anthropic evaluation in India, however, remains unconfirmed.

Company Calls for Binding AI Transparency Requirements

Alongside the model-access discussions, Anthropic is advocating binding rules that require AI companies to be transparent about their models. Mohanty said the requirements should apply to companies operating in India regardless of whether they are domestic or foreign.

The position raises a central question for AI governance: what information should developers be required to disclose so that regulators and independent evaluators can assess the risks associated with their systems?

Potential transparency frameworks can cover areas such as documented model capabilities, known limitations, evaluation methods, safety testing and incident reporting. The specific requirements advocated by Anthropic in this instance have not been detailed in a formal proposal in the reported remarks.

Mohanty also distinguished between meaningful evaluator access and access to model weights, which he described as a trade secret. This points to a debate over how external reviewers can scrutinise powerful AI systems without requiring developers to disclose every element of their proprietary technology.

For policymakers, the challenge is to determine what information is necessary for credible oversight, who should receive it and how sensitive commercial or security-related information should be protected.

India Reassesses Its AI Regulatory Approach

The discussions come as the Indian government considers whether the country needs a dedicated law for artificial intelligence.

India’s AI Governance Guidelines, released in November 2025, initially favoured using existing legal frameworks and governance mechanisms rather than immediately introducing a standalone AI statute. Since then, government statements have indicated growing interest in examining more direct regulatory measures.

MeitY has also signalled plans for a regulation-focused consultation paper addressing AI safety and governance. The precise legislative path and any future disclosure obligations remain subject to government decisions.

The distinction between existing requirements and proposed rules is important for companies developing or deploying AI in India. A call for binding transparency does not itself create a legal obligation. Any new requirements would need to be defined through the appropriate policy or legislative process.

For enterprises, changes in this area could eventually influence how AI suppliers document model risks, support independent assessments and communicate limitations to customers. The practical impact will depend on the final scope of any rules and how they apply across different AI systems and use cases.

Evaluating Advanced AI Beyond Content Moderation

Mohanty also argued that safety approaches designed primarily for online content moderation are insufficient for AI agents that can operate over extended periods, interact with other agents and take actions in real-world environments.

This distinction matters as AI products move beyond generating text and images towards systems that can use software tools, interact with digital services and perform multi-step tasks. Their risks may involve not only the content they produce but also the actions they take, the permissions they receive and the systems they can access.

Testing such systems may require evaluators to examine behaviour across sequences of actions, including whether an agent respects access controls, handles unexpected situations safely and avoids unintended consequences. These are broader questions than whether an individual response contains prohibited or misleading material.

Mohanty also emphasised the importance of ensuring that cyber defenders can access capable AI tools to identify and address vulnerabilities. This reflects a wider security debate about how advanced models can support defensive work while limiting the possibility of misuse.

The announcement did not set out a detailed testing protocol or establish new cybersecurity deployment rules for Indian organisations.

Independent Evaluation Remains a Governance Challenge

Anthropic’s discussions with MeitY also sit within a broader debate over how independent AI safety assessments should work. Evaluators need enough access to test systems meaningfully, while developers have incentives to protect customer information, proprietary technology and sensitive security details.

A credible framework must define the scope of access, protect evaluator independence and establish how findings can be reported. Transparency rules could help create consistent expectations across companies, but their effectiveness would depend on clear standards and practical enforcement.

For India, the next milestones are whether the discussions lead to a formal testing arrangement and how the government develops its wider regulatory framework.

For now, Anthropic’s position signals support for more structured AI oversight while preserving boundaries around proprietary model components. The outcome of the discussions could help clarify how India approaches the relationship between model access, safety evaluation and mandatory transparency as AI capabilities continue to advance.