The Lift Line
Zero Trust was built to keep out anyone who could not prove they belonged. It has no answer to an attacker that can.
Why This Editorial Matters for Your Exam
Cybersecurity answers in GS3 tend to be lists: CERT-In, the National Cyber Security Policy 2013, critical information infrastructure, a gesture at capacity-building. This editorial supplies the mechanism layer, why AI changes the character of the threat rather than its volume, which converts a list into an argument. The generative-versus-agentic distinction and the Zero Trust critique are both new enough that most candidates will not have them.
GS Paper 3: Basics of cyber security; challenges to internal security through communication networks; role of media and social networking sites in internal security challenges; awareness in the fields of IT and computers.
GS Paper 2: Important international institutions and groupings, and the gaps in global governance architecture.
| Concept | Meaning | Why it is testable |
|---|---|---|
| Generative AI | Systems that produce text, code or images in response to a prompt | The familiar category; useful as the contrast term |
| Agentic AI | Systems that pursue multi-step goals autonomously, taking actions without prompting at each step | Where the editorial locates the danger; the distinction is the examinable point |
| Zero Trust | A security model that grants no implicit trust and verifies every request regardless of network location | The current enterprise standard, and what the editorial argues is being structurally defeated |
| Zero-day vulnerability | A flaw unknown to the vendor, for which no patch exists | AI-assisted discovery inverts the discovery-to-patch relationship |
| Insider threat | Risk arising from an actor operating with legitimate access | The vector an autonomous agent with valid credentials occupies |
Background and Context
The author is M.K. Narayanan, former National Security Adviser and former Governor of West Bengal, writing in The Hindu on 4 August 2026.
The organising claim is that AI has crossed a threshold. So long as AI was an efficiency mechanism, adopted where it improved a process, its failures were contained to that process. Once AI becomes an infrastructure mandate, embedded in power dispatch, financial settlement, clinical decision support and telecommunications routing, an AI failure is an infrastructure failure. India is directly exposed here. CERT-In handled about 29.44 lakh incidents in 2025, up roughly 44 per cent on the 20.41 lakh of 2024, and the World Economic Forum’s Global Risks Report 2026 ranks cybersecurity as India’s single largest national risk, ahead of economic downturn, climate disaster and armed conflict. Meanwhile the country’s power grids, banking, health systems and telecom have been substantially digitised over a short period.
The Core Argument / Issue
The three multiplications
| Axis | What AI changes | Why the existing defence fails |
|---|---|---|
| Autonomy | Malware adapts to the defensive environment it encounters | Signature-based detection matches known patterns; self-modifying code has no stable pattern |
| Accessibility | Capability once requiring state resources becomes available to non-state actors | Threat models built around a small number of identifiable state adversaries misdescribe the threat population |
| Speed | Attack sequences compress below human decision cycles | Any defence with a human in the loop is, by construction, too slow |
The third is the one with the uncomfortable implication. If the attack completes faster than a human can respond, the only viable defence is also autonomous, which means delegating consequential security decisions to machines. That is a governance problem dressed as a technical one.
Why the Zero Trust critique is the sharpest claim
Zero Trust replaced the older perimeter model, in which anything inside the network was trusted. Its principle is that trust is never implicit and every request is verified regardless of origin. Against an external attacker attempting to appear internal, this works well.
An autonomous malicious agent presents a different problem. It operates with credentials that are genuinely valid, having been obtained through phishing, compromise or misconfiguration, and it can behave in ways statistically indistinguishable from an authorised user because it can model what authorised behaviour looks like. Verification succeeds, because there is nothing wrong with the credential. The attacker is not defeating the verification; it is satisfying it.
This is why the editorial’s framing matters, and it is also where the framing should be handled with care. Narayanan’s claim is not that Zero Trust is badly implemented or under-funded, but that its foundational assumption, that verification distinguishes the legitimate from the illegitimate, does not hold against an adversary that can be verified and is illegitimate.
The security literature agrees on the mechanism and is more divided on the conclusion. It broadly accepts that an autonomous agent occupies the insider-threat vector that credential checking cannot reach; the working consensus is that every autonomous agent should be treated as an insider threat, and the number of non-human identities in enterprise environments now vastly exceeds the number of human ones. But the prevailing prescription is to extend Zero Trust to non-human principals, through per-agent identity, short-lived and narrowly scoped credentials, runtime authorisation and continuous behavioural monitoring, rather than to abandon it. Narayanan’s own remedy, notably, is “Zero Trust 2.0”, which is an upgrade rather than a replacement.
A candidate should therefore present the strong form as the columnist’s position and the extension response as the field’s, because the difference between them is exactly the kind of distinction an examiner rewards.
The vulnerability-discovery inversion
Vulnerability management rests on an implicit race. Researchers and vendors find flaws and issue patches; attackers find flaws and exploit them. The system is tolerable because discovery is slow, expensive and requires scarce expertise.
AI systems capable of independently discovering zero-days remove the scarcity constraint. If discovery becomes cheap and fast while patching remains slow, because patching requires vendor action, testing, distribution and, in critical infrastructure, scheduled downtime, then the stock of exploitable unpatched vulnerabilities grows structurally rather than incidentally.
The governance vacuum
| Domain | Governing architecture |
|---|---|
| Nuclear weapons | NPT, CTBT, bilateral arms-control treaties, IAEA safeguards |
| Chemical weapons | Chemical Weapons Convention, with OPCW verification |
| Biological weapons | Biological Weapons Convention, without a verification protocol; negotiations on one collapsed in 2001 |
| AI-enabled offensive cyber operations | Thin, fragmented and largely non-binding on offensive use |
The accurate statement is not that nothing exists, but that what exists is thin and that India is outside all of it. That is a sharper and more useful point than a blanket vacuum claim.
| Instrument | Status | India |
|---|---|---|
| Budapest Convention on Cybercrime (2001) | Binding on parties; India has declined since 2001, on sovereignty and data-sharing grounds and because it was drafted without Indian participation | Not a party |
| UN Convention against Cybercrime (opened for signature at Hanoi, 25 October 2025) | Around 65 to 70 states signed on the opening day | Has not signed |
| Council of Europe Framework Convention on Artificial Intelligence, Human Rights, Democracy and the Rule of Law (adopted May 2024, opened for signature September 2024) | The first legally binding international AI treaty; signed by the United States, United Kingdom, European Union, Japan, Canada, Australia and others | Not a signatory |
| Lethal autonomous weapons systems | Informal meetings under the Convention on Certain Conventional Weapons since 2014, a Group of Governmental Experts since 2016; no binding instrument | Participant, no binding obligation |
| State behaviour in cyberspace | UN GGE norms of 2015, voluntary; the Open-Ended Working Group concluded in 2025 with a successor permanent mechanism | Participant, voluntary only |
None of these constrains AI-enabled offensive operations specifically, and the acceleration of the United States and China capability race is proceeding without one. Narayanan’s own formulation is that international law here is in a fledgling state without unified enforcement, which is more precise than saying it does not exist.
How to Think About This (Analytical Frame)
When a defensive system fails, distinguish failures of implementation from failures of assumption, because the two require entirely different responses. An implementation failure is met with more funding, better training, wider deployment. An assumption failure is met by redesigning the model, and no amount of investment in the existing model helps. The diagnostic question is: does the attack succeed despite the defence working correctly? If it does, the assumption is wrong. Apply this to fraud detection, border management, financial regulation and epidemiological surveillance alike; systems under pressure reliably respond to assumption failures by funding implementation, because that response is available and the other is not.
The Diagram in Words
Picture a building with a guard at every door who checks identification against a register. That is Zero Trust, and it is a genuine improvement on the older arrangement of one guard at the front gate and free movement inside. Now picture a visitor who holds a genuine pass, issued to a real employee, and who has studied enough of that employee’s routine to walk the corridors in a way nobody finds odd. Every guard checks the pass. Every check passes. The building’s security is functioning exactly as designed, and the intruder is inside. The problem is not that the guards are lazy or too few. It is that the building’s security was built on the belief that identification distinguishes friend from adversary, and that belief has stopped being true.
Way Forward
- Preserve meaningful human control over military decision-making, and press for international norms on autonomous weapons and AI-enabled operations in a domain that presently has none.
- Mandate pre-deployment safety testing for AI systems entering critical infrastructure, on the same logic that governs pharmaceutical and aviation certification.
- Build indigenous AI and semiconductor capability, on the reasoning that a country that imports its computational substrate cannot fully secure it.
- Fund cybersecurity research and skilled manpower, since the binding constraint on Indian cyber defence is people rather than policy documents.
- Harden critical infrastructure specifically, treating power grids, banking, health systems and telecom as a distinct protection category rather than as ordinary enterprise networks.
- Address algorithmic radicalisation as a security matter, since personalised feed distortion is a societal-security threat that sits outside the conventional cyber portfolio.
PYQ Linkage and Practice
UPSC has tested cyber security, critical information infrastructure, social media and internal security, and the challenges of emerging technologies across recent GS3 cycles. This editorial supplies the mechanism that connects the AI questions to the cyber questions, which most answers currently treat as separate syllabus entries.
Practice question: “The convergence of artificial intelligence and cyber capability has rendered prevailing defensive architectures obsolete at the level of assumption rather than implementation.” Examine this claim, and assess what institutional and international responses it implies for India. (250 words, 15 marks)
Interview angle: If an AI system can independently discover zero-day vulnerabilities, the same capability that lets an attacker find them lets a defender find them first. What determines which side that capability favours, and is there any policy lever that shifts the balance toward defence?
Sources: The Hindu, CERT-In, Ministry of Electronics and Information Technology
Source: The Double Helix: Why AI and Cyber Are No Longer Two Problems — Ujiyari.com | Free UPSC & State PCS Editorial Analysis
