Cybersecurity & Data Security

Overview

India's cybersecurity landscape is undergoing a structural transformation driven by AI—simultaneously becoming a more powerful defensive tool and a force multiplier for adversaries. With over 500 million financially underserved citizens entering digital systems 1, the stakes of getting security wrong are not merely commercial but civilizational. Threats have evolved from opportunistic attacks to industrial-scale, cross-border fraud operations that exploit weak inter-institutional coordination, legacy infrastructure, and a growing deepfake ecosystem that no reliable detection technology can yet contain 23. At the same time, a quiet but consequential deadline is approaching: quantum computing threatens to render current encryption obsolete, and most Indian organizations have not yet begun the migration 4. The decisions made in the next three to five years—on AI security standards, quantum-resistant cryptography, data sovereignty, and cross-border enforcement—will determine whether India's digital expansion is built on a foundation that can hold.

---

Key Insights

• Adversarial AI has outpaced defensive AI in one critical dimension—intent detection. Current systems excel at anomaly detection but cannot reliably identify social engineering before a transaction completes. Until behavioral and intent-based models mature, scammers will continue adapting faster than defenses can respond 3.

• The EN 304223 international standard for AI security is available now and freely accessible through ETSI. It covers all five lifecycle phases—design, development, deployment, maintenance, and end-of-life—and distinguishes mandatory baseline requirements from optional enhancements, making it usable by resource-constrained organizations without sacrificing rigor for security leaders who want to exceed the baseline 5.

• Post-quantum cryptography migration cannot wait for certainty on timelines. Organizations—particularly in critical infrastructure, finance, and healthcare—must immediately inventory where cryptography is deployed, classify data by sensitivity and lifespan, and begin hybrid encryption testing. A software-library-based migration path exists that avoids rip-and-replace disruption 4.

• Mule accounts, not individual scammers, are the operational lever of large-scale fraud. Disrupting the money-laundering infrastructure that moves stolen funds is more systemically effective than chasing individual perpetrators. This requires coordinated intelligence sharing across law enforcement, financial institutions, and telecom providers—not just better fraud models within a single bank 3.

• Security-by-design for agentic AI systems is urgently needed before, not after, wide deployment. Autonomous agents operating with delegated permissions, without transparency about their decision scope, create accountability voids that neither current regulation nor technical patches adequately address 61.

• India's observability infrastructure is insufficient for the security demands of AI at scale. Without domestic telemetry capabilities integrated across network and ML layers, operators cannot diagnose attacks or failures. Telemetry must be kept onshore—offshoring observability data is itself a sovereignty risk 7.

• Deepfakes are on a trajectory toward becoming a national security issue by 2029, with no reliable detection technology currently available. The combination of cheap social media amplification and increasingly convincing synthetic media creates conditions for coordinated disinformation at a scale that judicial and law enforcement systems are structurally unprepared to handle 2.

• The cyber-AI security divide will replicate the existing digital divide across sectors. Financially mature sectors like banking have measurably higher cybersecurity maturity than healthcare or agriculture. Without standard assessment frameworks and mandatory capacity-building programs, AI will widen this gap rather than close it 8.

• Boards must reframe AI security from a compliance checkbox to a quantifiable strategic risk. Reputation damage, service-provider dependency, and model reliability failures carry financial exposure that can be modeled and communicated to stakeholders—but most boards are not yet having this conversation in those terms 8.

---

Recurring Themes

• "Secure by design" cannot be retrofitted. Speakers across the standards 5, agentic AI 6, and infrastructure 7 sessions independently converged on the same principle: security architectures built after deployment are structurally weaker and exponentially more expensive to enforce. The implementation window for AI is narrower than it was for earlier digital technologies, and that window is closing.

• Human oversight remains non-negotiable, even as automation accelerates. Whether the context was autonomous financial agents 1, scam-detection systems 3, or AI infrastructure governance 6, every session that touched on automation drew the same boundary: humans must retain veto authority and must understand, at minimum, what categories of decisions are being made on their behalf.

• Individual institutional solutions are insufficient; coordination is the unsolved problem. Voice fraud 9, financial scams 3, deepfake disinformation 2, and post-quantum migration 4 all share a structural feature: the threat is cross-border and cross-institutional, while defenses remain siloed within organizations or national boundaries. Real-time intelligence sharing—across banks, telecoms, law enforcement, and nations—was identified as the critical missing layer.

• India's vendor dependency and data sovereignty exposure are underappreciated risks. Reliance on American API credits 9, offshore telemetry 7, and foreign-controlled cryptographic infrastructure 4 was flagged repeatedly as a systemic vulnerability that compounds every other security risk. Building domestic alternatives was framed not as protectionism but as operational necessity.

• Principles have been written; implementation is the actual challenge. UNESCO frameworks, RBI guidelines, and voluntary industry standards exist. The gap between their existence and their enforcement—through monitoring, auditing, and meaningful accountability mechanisms—was described across multiple sessions as the defining failure mode of the current moment 165.

---

Open Challenges & Tensions

• Real-time fraud prevention versus forensic accuracy is a genuine trade-off, not a false choice. The buildathon session surfaced a concrete disagreement: banking fraud systems operating at 3–5 second latency can flag high-risk calls if accuracy exceeds 95%, but achieving that accuracy threshold in real time, across noisy voice channels with caller ID spoofing, remains unsolved at production scale 93.

• The defensive AI asymmetry has no clean resolution. Defensive systems are constrained by privacy regulation, customer experience requirements, and legal liability. Offensive AI operates under none of these constraints. The honest conclusion from this structural asymmetry is that the goal cannot be eliminating fraud—it must be raising the cost and risk to attackers faster than they adapt—but this framing remains politically difficult to communicate publicly 3.

• Attribution and cross-border enforcement are structurally broken, and no one has a near-term fix. Mutual legal assistance treaties, Interpol cooperation, and financial intelligence sharing are all acknowledged as necessary but individually insufficient. The session on cybercrime explicitly accepted that some cybercriminals will evade prosecution indefinitely and recommended shifting resources toward victim protection and deterrence—a significant concession that policy frameworks have not yet absorbed 2.

• Post-quantum timelines are uncertain, but the cost of waiting is not symmetric. Organizations face a genuine planning dilemma: beginning expensive migration efforts before Cryptographically Relevant Quantum Computers (CRQCs) exist, or waiting for certainty and potentially finding critical data already compromised through "harvest now, decrypt later" attacks. The sessions did not resolve this tension but unanimously recommended acting before certainty 4.

• Domestic AI infrastructure development is essential but carries a 5–10 year honest timeline. The call for Indian-built LLMs, domestic telemetry, and sovereign cryptographic infrastructure is strategically sound, but the gap between current capability and that goal is substantial. In the interim, India's critical systems run on foreign infrastructure—a risk that neither the government nor the private sector has fully priced in 97.

---

Notable Examples

• RBI's Digital Payment Intelligence Platform was cited as a meaningful step toward national-level fraud coordination, but was explicitly assessed as necessary-but-insufficient without real-time cross-border intelligence sharing and enforcement mechanisms that match the international operational footprint of organized fraud rings 3.

• India's National Quantum Mission, including national test beds, PQC mandates, and democratized access programs, was highlighted as a model for proactive, vision-led technology governance—contrasted favorably against the reactive posture most nations took toward AI regulation 4.

• EN 304223, the new ETSI international AI security standard, was presented as immediately usable baseline infrastructure, developed through multistakeholder consultation spanning governments, industry, academia, and civil society across all continents—giving it practical credibility that unilateral government standards typically lack 5.

• Buildathon teams including Kartav, Walker Penguins, and Analytics were cited as concrete examples of the fragmentation problem in AI security tool development: each team independently built complementary capabilities (call analysis, speed, audio splicing detection) that judges argued should be integrated into a unified platform rather than deployed as competing solo solutions 9.

• UNESCO's Readiness Assessment Methodology (RAM) and ethical impact assessment tools were identified as existing, deployable infrastructure for translating AI governance principles into practice—specifically named as essential for financial inclusion AI systems serving the 500+ million underbanked population that faces heightened data bias risk 1.