New Delhi: Anthropic is discussing access to its artificial intelligence models with the Ministry of Electronics and Information Technology (MeitY) for testing by the IndiaAI Safety Institute. The developer of Claude also wants mandatory transparency requirements for Indian and foreign AI companies, its India policy head, Amlan Mohanty, said Thursday.
Speaking at the Trust and Safety India Festival in New Delhi, Mohanty said discussions were under way on making the models available for evaluation. His remarks indicated willingness to cooperate, rather than a completed agreement specifying access conditions or a testing timetable.
The proposed cooperation would build on an offer made by Anthropic’s chief executive, Dario Amodei, at the India-AI Impact Summit in February. Amodei had expressed interest in working with India on assessing security risks and studying AI’s economic consequences, identifying the country as a potential partner in addressing both challenges.
India’s safety institute was announced in January 2025 under the IndiaAI Mission, with a mandate to advance research suited to the country’s needs. According to the government’s announcement, it brings together academic institutions, businesses, startups and government bodies through a hub-and-spoke structure, with research grounded in India’s social, cultural and linguistic diversity.
The government’s India AI Governance Guidelines envisage the institute conducting safety research, evaluating systems and advising policymakers and industry. The framework combines technical safeguards, voluntary measures and existing legal obligations, while recommending that regulatory gaps be addressed through targeted changes to legislation.
Mohanty’s demand for enforceable disclosure requirements adds to the debate over how much oversight should depend on companies’ own commitments. He also argued that conventional content-moderation systems were inadequate for AI agents capable of sustained activity, interaction with other agents and action in the real world.
The transparency question extends to the independence of external assessments. A Reuters investigation published on Tuesday reported that Anthropic and OpenAI had sought substantial control over what outside evaluators could publish, citing concerns about proprietary information; Anthropic disputed the report’s characterization of all its discussions.
Anthropic has separately begun developing an arrangement under which evaluators work inside its laboratories. In its September 18 announcement of a partnership with Accenture, the company said Faculty, Accenture’s specialist AI business, would assess models, test safeguards and conduct adversarial exercises designed to expose weaknesses.
The company said embedded evaluators would receive access comparable to employees, enabling them to observe training, examine deployment decisions and speak directly to staff. However, it acknowledged that common standards governing access, publication and funding remained unsettled, and said it would initially pay directly for Accenture’s work while favouring pooled or government funding over the longer term.
Cybersecurity illustrates the competing considerations behind model access: capabilities that help defenders identify weaknesses can also assist attackers. Anthropic’s expanded Cyber Verification Program, announced on October 6, provides different levels of access for defensive work, authorized penetration testing and assessments of particularly sensitive systems.
The most permissive category is restricted to verified organizations authorized to test systems such as power grids, telecommunications networks and interbank transfer infrastructure. Anthropic says applicants for that category undergo detailed review in collaboration with the US government, demonstrating that access to advanced capabilities remains subject to security controls.
The discussions also coincide with Anthropic’s growing commercial presence in India, where it opened its Bengaluru office on February 16. The company identifies India as Claude.ai’s second-largest market and says it is developing locally relevant evaluations with partners, including tests involving Indian languages, agriculture and law.