Agentic AI may close fintech’s cyber defence gap

This article is authored by Soham Jagtap, senior research associate, The Dialogue.

India’s foundational Digital Public Infrastructure (DPI) comprising Jan Dhan Accounts, Aadhar and Mobile have revolutionised financial inclusion and welfare delivery by making digital transactions accessible. India is continually experiencing an upward surge in digital payments volume and value, most notably augmented by Unified Payments Interface (UPI). Digital transactions make up about 99.8% of all transactions by volume, with payment volumes having grown 38x and Compound Annual Growth Rate (CAGR) by 52.5% in volume and 13% in value in the past decade. Further, in 2025, India processed an average of 20 billion UPI transactions on average per month.

AI (Unsplash)

This scale represents a double-edged sword. Every transaction is a data point; every data point is an attack surface. The interconnected nature of the Banking, Financial Services, and Insurance (BFSI) sector and the sensitive nature of financial data that it secures presents a unique cybersecurity challenge. In recent years, Indian financial institutions faced a significant surge in cyber incidents, with the fintech sector emerging as a primary target. Artificial Intelligence (AI) has emerged as an accessible and increasingly affordable tool used by attackers to carry out sophisticated and pervasive identity-based attacks. Phishing attacks alone account for 25% of initial infection vectors, and the emergence of malicious Large Language Models (LLM) like WormGPT and FraudGPT has made them cheaper and more convincing than ever. A less-skilled bad actor can now generate targeted phishing emails, produce malware, and exploit vulnerabilities at scale, capabilities that once required significant technical expertise.

This has resulted in fintech cybersecurity (FinSec) becoming an asymmetric venture. Attackers need to exploit just one vulnerability in the system, immensely increasing the attack surface. On the other hand, security operations must account for every possible loophole on every security layer, constantly staying updated on emergent threats. Human vulnerability remains a critical risk in cybersecurity, particularly in the age of Gen AI. A Stanford study underscores this reality, finding that 88% of cybersecurity breaches are attributable to human error. Further, countermeasures require timely interventions which are impossible to carry out using only human capabilities. Addressing this asymmetry necessitates a techno-legal response: autonomous, agentic AI-driven defence systems governed by a regulatory architecture built for the threat landscape of today.

Agentic AI can help fintechs take autonomous, timely actions based on certain trigger events. Its capabilities can autonomously navigate diverse computing architectures and correlate complex attack vectors ranging from SQL to prompt injection at machine speed, which would otherwise require immaculate cooperation among humans with various Information Technology (IT) and security skills to pull off. A single, optimised agentic AI model can synthesise the capabilities of multiple security professionals and act swiftly and autonomously based on pre-ordained triggers, enabling real-time detection and neutralisation required to close the asymmetry gap.

The case for agentic AI in FinSec is compelling, but its integration introduces a paradox. The same democratisation of AI that makes autonomous defence necessary also means that a poorly governed AI security architecture can become a liability rather than an asset. Deploying agentic AI requires rigorously defining access controls, mapping permissions, and establishing accountability frameworks as prerequisites. Poor AI governance and access controls has vehemently plagued organisations, with 97% of them reporting AI-related security incidents that lacked proper AI access controls, while 63% had no AI governance policies to manage or prevent shadow AI. As fintechs move towards agentic AI, the question is no longer about building higher walls, but about ensuring the right roles, identities, and permissions are in place to protect high-value assets. Without this foundation, agentic AI is as likely to widen the vulnerability surface as it is to close it.

Agentic AI adoption in FinSec operates within a governance vacuum that neither organisations nor regulators have adequately addressed. As agentic AI systems make autonomous decisions, including blocking transactions, flagging accounts, and escalating incidents, clear accountability and auditability frameworks become non-negotiable. The Reserve Bank of India’s (RBI) FREE-AI Committee report gestures towards this through Recommendation 15, which directs regulated entities to draft AI policies covering governance structure, risk appetite, auditability, liability, etc. However, standardised policy templates remain a work in progress, with the broader ecosystem still grappling with the fundamentals of AI onboarding. In the interim, regular stress testing of AI-augmented security postures in controlled sandboxes is essential to simulate adversarial scenarios and prepare for unknown threats. Industry bodies have a role to play in supporting smaller entities through this process, as the FREE-AI Committee itself acknowledges. India would do well to adapt the NIST AI Risk Management Framework's “govern, map, measure, manage” structure to the Indian context, as a foundation for a standardised, responsible approach to AI adoption in FinSec. The RBI has already drawn on NIST's Cybersecurity Framework to move beyond compliance toward a broader governance and cyber resilience posture. Extending that thinking to AI risk management in finance is the natural next step.

AI security, much like traditional cybersecurity, must be approached as a layered discipline. Instead of focusing on securing the model in isolation, choosing the right model for the right context, rigorously testing it, and governing its behaviour through filters, guardrails, and accountability frameworks must be a priority. A chatbot deployed for customer engagement cannot be treated the same as an AI agent embedded within a financial system, healthcare workflow, or fraud engine. Each use case requires a different set of guardrails, filters and access controls. Security-by-design, continuous observability, and cross-functional oversight, from CIOs to privacy teams, must become non-negotiable. In an era where the cost of AI-enabled attacks is rapidly declining, resilience will belong to institutions that govern AI as a living system, not a one-time deployment. Over the next three to five years, as attacks grow more complex, AI-assisted first instance responses and security standards will have to be codified.

By the end of this year, the RBI and Securities and Exchange Board of India (SEBI) must aim to issue specific guidelines to concretise the recommendations elucidated in the FREE-AI committee report. A risk-based approach can be considered to categorise the usage of AI in Fintech and mandating proportionate compliances. In the next couple of years, CERT-In must consider updating its incident reporting framework to include AI-specific incident categories such as model manipulation, prompt injection and adversarial inputs. Lastly, in the next five years, regulatory bodies including the RBI, CERT-In and ministry of electronics and information technology (MeitY) must come together to codify minimum AI governance standards for fintechs, including mandatory red-teaming and sandbox testing.

Responsible experimentation today must harden into codified standards tomorrow. The aspirational nature of governance frameworks at the present moment must evolve and concretise over time to ensure that bad actors aren’t better equipped to use AI for attacks compared to the defensive capabilities of good actors.

This article is authored by Soham Jagtap, senior research associate, The Dialogue.