The AI Race Will Be Won By Countries That Can Recover, Not Just Innovate

The AI race will be won by countries that can recover, not just innovate. While the global focus often centers on AI models, chips, and data centre capacity, these factors pale in comparison to a country's ability to maintain functioning AI systems when faced with attacks or failures. Across critical sectors, AI is already well beyond its experimental stage.

Over 1,600 AI-enabled medical devices have received FDA approval, and an estimated 81% of North American utilities rely on AI-related technologies. In India, the eSanjeevani eHealth Assistance System supported 282 million consultations from April 2023 to November 2025, with 12 million aided by AI-recommended diagnoses.

However, this increasing integration of AI into public services and critical infrastructure creates both efficiency gains and heightened dependencies. When these dependencies fail, the consequences extend beyond IT teams to patients, customers, and citizens. Furthermore, attackers can exploit AI systems through various means, such as corrupting training data, manipulating inputs, or influencing system behavior.

Researchers demonstrated that as few as 250 malicious documents could introduce a backdoor into language models ranging from 600 million to 13 billion parameters.

Another risk dimension lies in the infrastructure connecting AI agents to real systems. A Securin Labs scan of 2,518 servers revealed that nearly one in four was deemed at critical risk. As AI agents gain the ability to take actions, the security implications become significantly more severe. Concentration of AI infrastructure among a few major cloud providers exacerbates this issue, with the three largest providers accounting for around 67% of global cloud infrastructure services spending.

To effectively address these challenges, governments and enterprises must adopt a proactive security approach from the outset. Critical questions, such as which data an AI system can access, where sensitive data and models will be hosted, who can alter them, and how to respond if a model produces unreliable output or becomes unavailable, need early consideration. Unfortunately, many organizations fail to adequately address these questions, treating security as an afterthought.

National AI strategies must also account for these complexities. Testing must keep pace with the rapid evolution of AI systems. Traditional penetration tests or annual audits are insufficient for continuously changing AI environments. Instead, continuous adversarial testing and red teaming under realistic attack conditions are essential for obtaining a more accurate assessment of system vulnerabilities.

Governments are recognizing this need, with the UK AI Security Institute publishing cyber evaluations of frontier AI models, the European Union's AI Act requiring adversarial testing for AI models with systemic risk, and India establishing an AI Safety Institute to oversee model evaluation, red teaming, and risk assessment.

However, developing AI safety is not merely about checklists and initial deployments. It requires ongoing discipline. Human factors play a crucial role, as voice cloning, synthetic identities, and deepfake videos succeed partly due to recipients' unawareness. Training employees to recognize suspicious activity, verify unusual requests, and know whom to alert remains a cost-effective security measure.

Yet, recovery planning often receives less attention, despite the significant costs associated with cyber incidents. For instance, following the Change Healthcare cyberattack, UnitedHealth recorded a $3.09 billion impact in 2024, and AIIMS Delhi faced two weeks of paper-based processes following a ransomware attack in 2022. Our research found that healthcare organizations experience ransomware roughly every ten hours, with one such incident occurring roughly every ten hours in the first quarter of 2026.