The next AI arms race will be over certainty, not intelligence

The recent restrictions on Anthropic's Fable 5 and Mythos 5 models highlight a broader tension emerging as frontier AI systems advance. What stands out is less the initial trigger than the US government's response: comprehensive, immediate and global in scope, with no public disclosure of underlying details. This step could set a precedent for major AI developers.

AI has reached an inflection point, enabling capabilities in cybersecurity, biology and scientific research. For the first time, this technology is being developed primarily by the private sector, with capabilities emerging in unpredictable ways. Frontier models demonstrate abilities not always anticipated during development, and the pace of advancement is becoming difficult to predict, even for their developers. Extensive testing, evaluation and red-teaming continue to improve safeguards. Yet, no provider can credibly offer absolute assurance.

National security governance has traditionally relied on a deterministic view of the world: threats are classified, risks are modelled, actors identified and intentions inferred, enabling policy instruments grounded in cause and effect. Existing frameworks depend on attribution, accountability, proportionality, deterrence, verification and jurisdiction, and were designed for human actors operating within stable boundaries.


Advancing AI introduces new risks. It changes the nature of risk. As agency becomes distributed, decisions are increasingly automated, explanations are system-generated, and consequences cascade beyond immediate visibility, weakening key assumptions of contemporary governance. The challenge isn't only regulation but comprehension.

The issue's no longer solely what a model can do. It is what governments fear they cannot predict. Preserving strategic advantage becomes central to decision-making, particularly in environments where uncertainty itself is treated as risk.

Any incident of compromise may be treated as a national security concern. The Mythos 5 restriction illustrates this dynamic: uncertainty, rather than scale of harm, becomes the primary driver of response. This emerges as geopolitical competition and tech protectionism reshape the global order. Its implications extend beyond any single company or country. Conditions enabling frontier AI - global talent mobility, open scientific exchange, distributed experimentation, infrastructure access and rapid diffusion - are also those that governments may seek to restrict when faced with capabilities they struggle to interpret and control.

India's position in this emerging dynamic is direct. Access to models is critical for leveraging potential from AI for the economy. The natural response is an accelerated push toward sovereign capability: indigenous models, domestic compute infrastructure and greater self-reliance - not primarily for commercial advantage but as insurance against strategic dependence and unilateral denial.

Yet, sovereignty pursued defensively introduces tension. The same conditions that support AI progress remain essential for participation in it. India has made significant strides in the global AI ecosystem, with AI data centre capacity potentially reaching 10-15 GW over the next 5 yrs. Pursued in isolation, defensive sovereignty risks undermining these ambitions.

India needs to double down on developing foundational models. MeitY's support for 20 model-development initiatives, with 5 released, is welcome. These models must be rapidly strengthened across domains. New approaches such as world models, with potential capability leaps, should be explored intensively.

Access to models, and assurances around their safety, security and governance, will have to be central in India's technology diplomacy. Investment in data centre capabilities in India will be a supporting pillar of this strategy. Promise of uninterrupted access to models should be embedded in policy positions. Equally, developing the domestic market will be critical for the success of Indian models and for strengthening India's position in the geopolitics of AI.

For companies planning to leverage frontier models and secure their codebases, digital footprints and supply chains, recent restrictions and associated uncertainty have been a shock. Security hygiene, therefore, should be elevated to match the risks posed by increasingly powerful AI systems.

A trimodal strategic approach should be adopted, with focus on immediate measures, developments expected in the next 6-12 mths, and structural capabilities required for AI advancement over the next 2-3 yrs.

India faces a challenge that extends beyond capability adoption. Strengthening domestic capacity while integrating with global AI ecosystems is one of the most critical policy objectives. The institutional question isn't whether to respond to this condition, but on what terms.

The central challenge of AI governance may not be controlling intelligence itself, but governing uncertainty as a persistent structural condition of technological progress.

(The writer is CEO, Data SecurityCouncil of India)