FATF AI and Deepfakes Horizon Scan 2026: What India Banks Must Do Under PMLA and FIU-IND

FATF's January 2026 Horizon Scan is the first dedicated FATF document to address AI as both a financial crime enabler and a compliance tool. It names three specific risk typologies that Indian banks must address: deepfake-enabled KYC bypass attacking V-CIP (Video-based Customer Identification Process) systems, AI-generated synthetic identities defeating liveness checks and document verification, and autonomous AI agents orchestrating transaction layering in ways that bypass rules-based AML monitoring. As a FATF member, India's obligations under FATF Recommendation 15 (new technologies) flow through FIU-IND — the Financial Intelligence Unit of India — and directly affect every reporting entity under the Prevention of Money Laundering Act (PMLA), 2002.

This page is distinct from Zytra's general AI AML governance page. It focuses specifically on the FATF Horizon Scan document published January 29, 2026 and what the three named risk typologies mean for India bank compliance under PMLA Sections 12 and 12AA, FIU-IND STR/CTR quality requirements, and RBI's V-CIP (Video-based Customer Identification Process) under the KYC Master Directions.

The Three Risk Typologies — What FATF Named

Risk Typology 1: Deepfake KYC Bypass

FATF's January 2026 Horizon Scan identifies deepfake technology as an acute threat to video-based customer identification. For Indian banks, the specific vulnerability is the V-CIP (Video-based Customer Identification Process) framework under RBI's KYC Master Directions. V-CIP requires a live video interaction for remote KYC — but advanced deepfake technology can now convincingly spoof a live video feed in real time.

What FATF requires banks to address:

• Liveness detection that goes beyond simple video checks — behavioral biometrics, depth sensing, challenge-response tests that are computationally difficult to spoof in real time
• Multi-modal identity verification: cross-referencing video identity against document verification, CKYC registry data, and behavioral signals — no single verification channel is sufficient
• Deepfake detection model validation: any AI model deployed for liveness detection or deepfake screening must be validated against current-generation deepfake capabilities, not capabilities as of the model's original training date

Risk Typology 2: AI-Generated Synthetic Identities

AI can now generate complete synthetic identities — name, address, employment history, digital footprint — that defeat document verification systems not designed to detect AI-generated inputs. The specific FATF finding: large language models combined with generative AI can produce KYC documentation (salary slips, utility bills, bank statements) that bypass optical character recognition and document verification systems trained on human-created documents.

India-specific risk: Aadhaar verification via OTP is a primary KYC channel — but Aadhaar-seeded KYC for an identity obtained through a compromised Aadhaar number is vulnerable to this attack. FATF specifically calls for document verification systems to be tested against AI-generated document types.

FIU-IND reporting implication: Suspicious Transaction Reports (STRs) filed for synthetic identity-related activity require documentation of the detection method. If the detection was made by an AI model, the STR must reflect that the model, not a human officer, identified the suspicion — and the model's basis for suspicion must be documentable.

Risk Typology 3: Autonomous AI Agent-Orchestrated Layering

FATF's Horizon Scan identifies a new class of money laundering risk: autonomous AI agents that execute transaction layering strategies across multiple accounts, jurisdictions, and asset classes faster than any human-supervised system can monitor. These agents can adapt their behavior in response to AML monitoring patterns — a technique FATF calls "adversarial AML evasion."

For India banks, this risk intersects with:

• UPI's high-velocity transaction infrastructure — an AI agent can execute hundreds of layering transactions per minute through UPI
• NPCI UAP (Unified Agent Protocol) — AI agents authorized to transact on UPI up to ₹15,000/month could in theory be repurposed for AI-orchestrated layering if the authorization scope is inadequately governed
• Rules-based AML monitoring limitations: India banks using threshold-based AML monitoring (fixed Rs. thresholds for STR generation) are specifically vulnerable to AI agents designed to stay below thresholds across large numbers of accounts

FATF Recommendation 15 — India's Compliance Obligation

FATF Recommendation 15 requires that countries and financial institutions identify and assess ML/TF risks associated with new and developing technologies. As a FATF member since 2010, India's compliance with Recommendation 15 flows through:

1. FIU-IND guidance: FIU-IND issues AML guidance to reporting entities under PMLA. The January 2026 FATF Horizon Scan is a FATF-level trigger for FIU-IND to update guidance on AI-specific AML risks. Banks should monitor FIU-IND for updated guidance and prepare compliance programs in advance.
2. PMLA Section 12 (record maintenance): Banks must maintain records of verification procedures used. If AI is used for any part of KYC, the verification record must include what AI system was used, what it verified, and its confidence/output — not just the conclusion.
3. PMLA Section 12AA (verification): The verification required under Section 12AA must be adequate to detect synthetic identities and deepfake presentations. "We used V-CIP" is not adequate documentation if the V-CIP system has not been validated against deepfake threats as of January 2026.

Connecting FATF Horizon Scan to RBI MRMF June 2026

The RBI Model Risk Management Framework (MRMF, June 2026) requires Board Risk Committee sign-off for high-risk AI models. FATF's January 2026 Horizon Scan creates a basis for classifying AML AI models — liveness detection, document verification, transaction monitoring — as high-risk under RBI MRMF, because:

1. They directly affect financial crime compliance obligations (PMLA, FATF)
2. Their failure creates regulatory liability, not just operational loss
3. They are subject to adversarial attack (FATF specifically names adversarial AML evasion as a risk)

Under RBI MRMF, a high-risk AI model requires: independent validation by a function separate from the model owner, documentation of adversarial robustness (resistance to intentional manipulation), and Board Risk Committee-level visibility including residual risk scores.

Five Practical Steps for India Bank AML AI Governance

1. Validate V-CIP AI against current deepfake capabilities (not training-time capabilities). Commission an adversarial test of your liveness detection and deepfake screening models against current-generation synthetic video. Document results in the model's validation record.
2. Test document verification AI against AI-generated document types. Current document verification models trained before 2024 may not detect AI-generated salary slips, utility bills, or bank statements. Retroactive testing is required.
3. Review NPCI UAP authorization scopes for AI agent accounts. For banks issuing UPI access to AI agents under the NPCI UAP framework, verify that authorization scopes cannot be repurposed for adversarial layering. Document kill-switch triggers.
4. Update STR documentation standards for AI-flagged suspicion. When an AI model generates the STR trigger, the filing must document the model's basis — not just the transaction description. Work with compliance counsel to establish standard language.
5. Establish Board Risk Committee visibility for AML AI models. Under RBI MRMF, classify liveness detection, document verification, and transaction monitoring AI as high-risk. Present residual risk scores to the Board Risk Committee with the next AML governance report.

Frequently Asked Questions