India’s AI Safety Institute must walk a fine line between safety and security at a time when global counterparts are shifting decisively towards security-first approaches.
Earlier this year, India announced that it would set up its own Artificial Intelligence Safety Institute (AISI). India’s AISI now faces the delicate challenge of walking a middle ground between the increasingly divergent conceptions of AI safety and security. Particularly over the course of 2025, there has been a strategic shift, with institutes in countries such as the United Kingdom (UK) and the United States (US) moving away from “safety” and toward “security”.
This article examines the global conceptions of AISIs and some of the recent security-oriented strategic shifts that AISIs have seen. It then maps India’s ongoing journey toward safe and responsible AI adoption and highlights how India’s AISI must chart its own unique path forward.
The foundation for international AI safety cooperation was laid in November 2023, when the Bletchley Declaration, following the UK AI Safety Summit, called on countries to make AI safety a priority, affirming that "AI should be designed, developed, deployed, and used in a manner that is safe". This declaration marked the beginning of a coordinated global effort to address AI risks.
Building on this momentum, the UK's Frontier AI Taskforce, which was meant to evaluate the risks associated with advanced AI models, would evolve into the world's first AI Safety Institute. The US would follow suit, with an AISI being housed within the National Institute of Standards and Technology to facilitate "rigorous assessment of AI risk". As this movement gained traction across the Pacific, Japan launched the J-AISI under its Information-technology Promotion Agency, with the Institute meant to collaborate with a number of ministries and departments in a "cross-government effort".
2025 has seen a significant shift in policy thinking, with some countries moving away from the "safety" tag when it comes to AI, partly out of trepidation that excessive regulation might inhibit growth.
The AI Seoul Summit in May 2024 saw the inception of an international network of AISIs, with several countries demonstrating interest in setting up their own institutes. This international network would convene for the first time in San Francisco in November 2024, with the US, Australia, and South Korea committing US$11 million towards AI safety research.
These AISIs perform a number of functions. The most important of these is evaluating advanced AI models, with the UK AISI conducting both pre- and post-deployment evaluations of model risks. They have also led to improved coordination and information sharing between a wide range of stakeholders. This extends to deliberating best practices on a global scale, thanks to the international network.
The conceptual foundation of AISIs was strongly tied to safety. This specifically meant ensuring that frontier AI models do not diverge from their objectives or produce negative side effects. The idea was to mitigate the challenges of bias, transparency, and accountability in complex algorithms in critical areas. However, 2025 has seen a significant shift in policy thinking, with some countries moving away from the "safety" tag when it comes to AI, partly out of trepidation that excessive regulation might inhibit growth. After all, increasing compliance requirements for frontier models increases the red tape that companies need to navigate, potentially deterring their plans for AI development. The result has been a growing emphasis on AI “security”, which seeks to combat targeted attacks on AI systems by external adversaries, often tied to national security.
This change in approach became evident at major international forums. In February 2025, the US and the UK did not sign the final declaration of the Paris AI summit, which called for "ethical and safe" AI development, with US Vice President JD Vance stating that "excessive regulation” could “kill a transformative industry".
The transformation was perhaps most striking in the UK, with the country rebranding its AI Safety Institute as the "UK AI Security Institute", with its focus shifting to national security and "protecting citizens" from crime. Similarly, while the Biden administration focused on AI safety, the emphasis has shifted to security under the Trump administration.
The result has been a growing emphasis on AI “security”, which seeks to combat targeted attacks on AI systems by external adversaries, often tied to national security.
This was evident when the US AISI was transformed into the Center for AI Standards and Innovation (CAISI) in June 2025, aiming to play two simultaneous roles. CAISI is mandated to ensure rapid AI innovation while aligning it with national security standards and objectives. Rather than focusing on the internal risks posed by models, both the UK AI Security Institute and the USA's CAISI are now geared towards risk assessment and mitigation of threats posed by the use of AI by other actors.
In January 2025, just a couple of weeks prior to New Delhi co-hosting the Paris AI Summit, it was announced that India would set up its own AI Safety Institute, incubated under the IndiaAI mission. This is understandable, since India has consistently affirmed a commitment to ensuring safe and responsible AI development and adoption. The National Strategy for Artificial Intelligence of 2018 identified “Safe and Trusted AI” as one of its key pillars. The Ministry of Electronics and Information Technology (MEITy) set up various committees to assess AI capabilities, with Committee-D specifically highlighting ethical issues and calling for AI systems “to meet a high degree of safety standards”. In 2021, an Approach Document for Responsible AI was released, developed collaboratively by NITI Aayog, multilateral forums, and private sector organisations. It highlighted key principles for responsible AI, relevant to both public and private stakeholders. Industry players have also been involved in the AI safety discourse through other means, such as by contributing to the NASSCOM Responsible AI Resource Kit. In 2023, MeitY called for organisations to send in research proposals related to Responsible AI, with 10 such projects receiving grants.
India's AISI has a unique structural design. It was conceived to run virtually, rather than at a specific location. MEITy led the process of setting it up, with the core aim of facilitating collaboration between researchers from different institutions across the country, as well as private sector partners, thereby creating a common platform.
On the safety side, the AISI must consider ways to institutionalise accountability, particularly for high-risk AI use cases. For instance, applications under consideration or implementation in India include predictive policing and credit scoring, both of which pose direct risks to citizens if trained on biased data.
The institute's collaborative framework became more concrete with subsequent developments. In May 2025, the AISI sent out a call for partner institutions, including academic institutions, startups, companies, and research and development (R&D) institutions, to help drive the AISI's mandate of indigenous R&D and developing India-specific AI governance practices. The idea is that each partner institution will maintain an "IndiaAI Safety Institute Cell", to which the AISI will allocate work depending on its specific expertise.
For India’s AISI, it is critical to strike an optimal balance between the safety and security components that many of its contemporary institutions are grappling with. After all, external and internal threats are not mutually exclusive, rendering sustainable safeguards necessary for both.
On the safety side, the AISI must consider ways to institutionalise accountability, particularly for high-risk AI use cases. For instance, applications under consideration or implementation in India include predictive policing and credit scoring, both of which pose direct risks to citizens if trained on biased data. To ensure algorithms are robust and not harmful, the AISI should prioritise setting up an independent oversight body, with clear funding and a skilled staff, empowered to ensure model accountability in high-risk use cases. The mandate of this body should allow it to examine data collection, cleaning, and storage practices in these domains. To make public sector deployment of AI more transparent, the AISI should also consider setting up a readily available and detailed digital registry. This should provide a comprehensive overview of all the major AI systems being deployed by public sector bodies, with easy-to-understand explanations of their technical applications and precise uses. Additional information, such as deployment status and the states where the model is implemented, would be useful, particularly if paired with a legible and easily accessible front-end interface. As similar efforts with AI registries in the Netherlands and Finland have demonstrated, such measures can help democratise AI safety.
AI security cannot be ignored either, particularly in the aftermath of recent hostilities with Pakistan, which deployed AI-driven disinformation and cyberattack campaigns.
AI security cannot be ignored either, particularly in the aftermath of recent hostilities with Pakistan, which deployed AI-driven disinformation and cyberattack campaigns. Emerging capabilities such as AI-enabled edge computing infrastructure will also become increasingly essential for security. In this context, the AISI provides an ideal platform to coordinate dialogue between industry, academia, and the military to develop a strategic roadmap for R&D into AI applications for national security. Organising an annual closed-door conference bringing together these stakeholders, with the end goal of such roadmap recommendations, would position the AISI as a central pillar of AI development in India.
While globally, countries face a critical choice between safety and security, India’s AISI possesses the unique opportunity to tackle both. The future of AI poses challenges both in terms of systemic side effects and deliberate attacks. The AISI can equip India with the capability to deal with both.
Dhruv Banerjee is a Research Intern with the Observer Research Foundation.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.
Dhruv Banerjee is a Research Intern with the Observer Research Foundation. ...
Read More +
PREV
