Chapter 31 · Cyber Risk, Data Protection and AI Governance
Part V changes the question. Parts I to IV asked what the instruments are and who regulates them; this part asks how the system keeps them safe, fair and open. Three regimes now meet inside every fintech: the cyber incident clock, the data-protection consent architecture, and the governance of models that decide who gets credit. This chapter teaches all three — through one Saturday-afternoon breach, three different deadlines, and the RBI's Seven Sutras for AI that nobody in Indian finance can now plead ignorance of.
🎯 Learning Objectives
- Map the three regimes that govern a fintech incident: cyber reporting, data protection, model risk
- Apply the incident clocks precisely — CERT-In's six hours, DPDP's without-delay and seventy-two-hour duties
- State the DPDP Rules 2025 architecture: consent, Consent Managers, security safeguards, erasure — and the phasing
- Define model risk and evaluate it against the FREE-AI framework's Seven Sutras
- Assemble the governance stack for a digital lender that uses AI and suffers a breach
- Run the AI governance checklist on any live product
1. The Perimeter Shifts
Everything this book has built — wallets, rails, credit engines, tokens — runs on three inputs: connectivity, data and models. Each input carries its own failure mode. A system can be attacked (cyber risk), it can mishandle what it holds (data risk), and it can decide wrongly at machine scale (model risk). For two decades Indian financial regulation treated these separately: an IT circular here, a privacy bill there, model governance as an internal bank matter. They have now converged — one regulator issues the incident rules (CERT-In's directions), another anchors the data regime's sectoral overlap (the RBI), and the RBI again has written the finance sector's AI doctrine. The organising skill of this chapter is holding all three in one frame, because in a real incident they arrive together, with different clocks.
2. Cyber Risk and the Incident Clocks
Fintech's attack surface is the book's earlier chapters read by an adversary. The phone is the branch (Chapter 10); social engineering beats cryptography every time (Chapter 14's fraud stack — the scammer's opening move is trust, not code); APIs connect everyone, including attackers; and lending apps hold exactly the data — contacts, messages, financial behaviour — whose collection Chapter 16's scandals made notorious. The regulatory response runs on two rails. CERT-In's April 2022 Directions require every entity to report specified cyber incidents — unauthorised access, data leaks, ransomware, spoofing — to the national CERT within six hours of noticing, around the clock. And the RBI's cyber security framework for regulated entities requires a board-approved policy, a dedicated chief information security officer, and prompt incident reporting to the Reserve Bank, with banks' timelines aligned to the CERT-In clock. A fintech with a bank partner therefore reports the same event to two masters on the same night.
What the clocks actually test is organisational honesty. Six hours from noticing — not from confirming, not from fixing — is a demand that firms report while frightened, incomplete and mid-incident. The industry's natural instinct is the opposite: investigate quietly, contain, then disclose a polished story. The rule exists because that instinct left regulators blind; the tension between the two never goes away, and the hands-on makes you live it. Note also what the incident anatomy implies for the other two regimes: the moment "unauthorised access to personal data" enters the incident log, this stops being only a cyber event — it becomes a data-protection event with its own duties, and if the compromised system feeds an AI model, a governance event too.
3. The DPDP Regime
The Digital Personal Data Protection Act, 2023 became fully real on 13 November 2025, when the government notified the DPDP Rules, 2025 and commenced the Act in phases. The architecture in one paragraph: a Data Fiduciary (the fintech) may process a Data Principal's (the customer's) personal data only for a consented, itemised purpose — with a standalone, plain-language notice; consent withdrawable as easily as given; data protected by reasonable security safeguards (encryption, masking or tokenisation, access control, logging with one-year retention, backups); erased when the purpose is served; children's data behind verifiable parental consent with tracking and targeted advertising barred; and breaches reported upward. Consent Managers — Indian companies, registered with the Board, interoperable — give individuals one dashboard for their permissions, an idea this book first met as UPI's consent architecture in Chapter 9, now generalised to all personal data.
The phasing is the exam question hiding in the rules. Immediate (from 13 November 2025): the institutional spine — the Data Protection Board, the government's rule-making powers. One year out (13 November 2026): Consent Manager registration. Eighteen months out (13 May 2027): everything operational — notices and consent, security safeguards, breach intimation, erasure, Data Principals' rights, penalties. So as this chapter is written, a lending app's notice duties are law-in-waiting, while CERT-In's six hours applies today — the trap the hands-on springs. And the breach duty itself is two-tiered when it arrives: affected users informed without delay, describing the breach, its consequences and the steps taken; the Board informed without delay initially, then with a detailed report — facts, causes, mitigation, findings, and proof of the user notifications — within seventy-two hours, extendable only on written request.
4. Model Risk and FREE-AI
A model is a decision rule learned from data. Model risk is everything that follows when the rule is wrong, skewed or opaque — and it scales with automation: a biased human officer misprices one loan; a biased model misprices a million, invisibly, with a confidence score attached. Chapter 16 met the raw material (alternative data signals), Chapter 18 the engines (OCEN, ULI), and Chapter 17 the conduct duties; what none of them supplied was the sector-wide doctrine. The RBI supplied it on 13 August 2025: the report of the Committee on the Framework for Responsible and Ethical Enablement of Artificial Intelligence — FREE-AI — an eight-member committee chaired by Dr Pushpak Bhattacharyya of IIT Bombay, 103 pages, twenty-six recommendations on six pillars.
The framework's spine is the Seven Sutras, and they are worth knowing verbatim, because they are already the vocabulary of Indian financial AI governance: Trust is the foundation; People first; Innovation over restraint; Fairness and equity; Accountability; Understandable by design; Safety, resilience and sustainability. Two design choices make the document distinctively Indian and distinctively RBI. First, its posture: innovation over restraint sits inside the principles themselves — the committee treats enabling innovation and mitigating its risks as complementary, not competing, a deliberate contrast with prohibition-first regimes. Second, its location of responsibility: boards of regulated entities own the AI policy — model risk is a governance duty, not a data science hobby. Run the sutras against a lending model and each becomes a test: Fairness and equity — does the feature set encode proxies for protected groups (a pincode is caste's neighbour)? Understandable by design — can the firm state to a rejected applicant, in the Digital Lending Directions' idiom, the principal reason for rejection? Accountability — when the model errs at scale, is there a named human owner, or only a vendor contract? This is the analytical move the session's checklist trains.
5. The Governance Stack
Stack the three regimes on one firm — a digital lender with an AI underwriting model — and see how they interlock. Before anything happens: the board owns a cyber policy (RBI), a data-protection compliance build aimed at May 2027 (DPDP), and an AI policy naming model owners and risk appetite (FREE-AI). At product design: the notice itemises every data field, including the model's features; consent is purpose-bound; the model's training data passes the fairness sutra; rejection-reason logic is built to be stateable, not discovered later. On incident day: the clocks fire together — CERT-In within six hours of noticing; the RBI through the entity's supervisory channel; and, once May 2027 arrives, users without delay and the Data Protection Board in detail within seventy-two. After: remediation feeds back into all three policies — the incident post-mortem is also a data-governance audit and, if training data was exfiltrated or poisoned, a model-risk review.
The stack's deepest lesson is that the three regimes are not three compliance costs; they are one trust system seen from three angles. A firm that itemises consent honestly has a smaller breach surface (it holds less), a cleaner incident story (it can tell users exactly what leaked), and a more defensible model (its features are lawful). Chapter 2 of this book argued that India's fintech runs on public rails; Part V's argument is that it survives on public trust, and the trust machinery is now written down. What remains — Chapter 32's subject — is who is still outside the machinery's reach.
6. The Checklist Method
The session's deliverable: a ten-point AI governance checklist that any student can run against any live product — and that maps every check to its authority: the sutras, the DPDP Rules, or the incident regime.
- Inventory the model. What decisions does it make, and who is affected at what scale?
- Name the owner. A named executive, accountable to the board — FREE-AI's accountability sutra.
- Audit the features. Every input field: consented? itemised? a proxy for a protected group? — DPDP notice duty plus the fairness sutra.
- Test the rejects. Can the firm state the principal reason, in plain language, to any declined applicant? — understandable by design, and Chapter 17's conduct law.
- Measure the skew. Outcome rates across groups the model never sees but society does.
- Check the fallback. Human review for edge cases and complaints; People first is operational, not decorative.
- Verify the data map. What is held, where, how encrypted, how long retained, when erased.
- Rehearse the clocks. Who calls CERT-In at 2 a.m.? Who drafts the user notification? Is the seventy-two-hour report template pre-written?
- Probe the vendor chain. Model, cloud, LSP: who processes what, under what contract clause — Chapter 17's LSP accountability applied to AI vendors.
- Report to the board. Model performance, incidents and complaints on a recurring agenda — the governance duty that makes the other nine real.
🛠 Hands-On: AI Governance Checklist
It is Saturday, 4:20 p.m., November 2026. LoanKart, a digital lender running an ML underwriting model, discovers that a misconfigured server exposed a database of two lakh customer records — names, PANs, loan repayments, and the model's raw features, including device model and SMS-derived spending signals. The security team believes the data was accessed by an unauthorised party ten days ago. Separately, a rejected applicant has filed a complaint alleging that the model rarely approves loans to her pincode cluster.
Steps
- Run the clocks. From discovery at 4:20 p.m. Saturday: who must be told, by when, in what form — CERT-In, the RBI, the partner bank, affected users, the Data Protection Board? State each deadline as a date-time, and flag any duty that is not yet in force.
- The consent audit. LoanKart's sign-up consent read: "to offer you credit and related services, we may collect information from your device and messages." Audit this against the DPDP notice standard — itemised data, specified purpose, withdrawal ease. Verdict?
- The model review. Apply the Seven Sutras to the pincode complaint: which sutra does each of the following engage — pincode as a feature, no rejection-reason logic, model owned by a vendor, board never briefed?
- The checklist verdict. Score LoanKart on the ten-point checklist — pass, partial, fail — and write the three remediation items you would put to its board on Monday.
View Solution / Walkthrough
1. The clocks. CERT-In: the exposure is a reportable cyber incident; report within six hours of noticing — by approximately 10:20 p.m. Saturday — with the ten-day-old access disclosed as known. The RBI: as a regulated-entity-linked LSP ecosystem event, report through the partner bank's supervisory channel immediately, per the entity's board-approved incident policy. Affected users and the Data Protection Board: the trap — DPDP Rule 7's without-delay user intimation and seventy-two-hour Board report are not in force in November 2026; the operational duties commence 13 May 2027. The defensible answer says both: no statutory DPDP duty yet, but a prudent entity notifies users anyway (contractual duty, RBI conduct expectations, and the reputational arithmetic of a ten-day-old leak), and files the CERT-In report that is due. The phrase "by when" earns its marks only with the phasing stated.
2. The consent audit — verdict: fail. The notice bundles ("we may collect information from your device and messages") where the Rule requires itemised description — each data category named; "credit and related services" where it requires specified purposes; nothing on withdrawal comparable to giving, on rights, or on complaint to the Board. And the model's training use of SMS signals is a purpose distinct from "offering credit" that a compliant notice would have to state separately. Remediation is not a reworded banner; it is a re-consent: itemise, explain, re-permission — or delete and retrain.
3. The sutras. Pincode as a feature: Fairness and equity — pincode is a geography-of-privilege proxy; test outcome rates across clusters and drop or justify the feature. No rejection-reason logic: Understandable by design — and a conduct breach under the Digital Lending Directions' disclosure duties. Vendor-owned model: Accountability — the regulated entity owns outcomes it cannot explain; FREE-AI and Chapter 17's LSP logic agree: you can outsource the model, not the responsibility. Board never briefed: the governance failure that underlies all the others — the AI policy FREE-AI assigns to the board was never written.
4. Checklist verdict (typical scoring). Fail: 3 (features unaudited, unconsented), 4 (no rejection reasons), 8 (no rehearsed clocks — a Saturday discovery with no template), 10 (no board reporting). Partial: 1, 2, 5, 7 (a data map exists but retention and erasure are unmanaged), 9 (vendor contract silent on breach cooperation). Monday's board items: (i) incident report filed and user notification issued voluntarily, with counsel's phasing memo; (ii) pincode feature suspended pending a fairness audit, rejection-reason logic commissioned; (iii) a FREE-AI-aligned AI policy with named model owner and quarterly board reporting — the three items that turn this Saturday from a crisis into the last ungoverned one.
Key Takeaways
Three regimes, one incident: the cyber clock, the data duties and the model-governance questions arrive together, on different deadlines.
Six hours is CERT-In's cyber clock, in force since 2022; DPDP's without-delay and seventy-two-hour breach duties start 13 May 2027 — knowing which clock is running is the test.
The DPDP architecture: itemised consent, Indian-registered Consent Managers, security safeguards, erasure, children protected — the Chapter 16 scandals' tombstone.
FREE-AI's Seven Sutras — trust, people first, innovation over restraint, fairness, accountability, understandable by design, safety and sustainability — with boards owning the AI policy.
The three regimes are one trust system: honest consent shrinks the breach surface, cleans the incident story and defends the model — governance is a moat, not a tax.
Test Your Understanding
Choose one best answer per question. The LoanKart scenario and the phase dates are fair data. Aim for 24/30 before moving to Chapter 32.
1. The three failure modes this chapter frames:
2. The six-hour reporting rule belongs to:
3. CERT-In's clock starts at:
4. The RBI's cyber framework for regulated entities requires all EXCEPT:
5. The DPDP Rules, 2025 were notified on:
6. The operational DPDP duties (notices, security, breach, erasure) commence:
7. The Consent Manager registration framework goes live at:
8. DPDP Consent Managers must be:
9. Under Rule 7, the detailed report to the Board is due:
10. A compliant DPDP notice is:
11. For children's data, the DPDP regime requires:
12. FREE-AI expands to:
13. The FREE-AI committee was chaired by:
14. Which is NOT one of the Seven Sutras?
15. The FREE-AI posture that distinguishes it from prohibition-first regimes:
16. FREE-AI locates AI responsibility with:
17. Model risk's defining property:
18. The sutra engaged by a pincode feature that correlates with caste or religion:
19. No rejection-reason logic in a lending model fails which sutra — and which earlier chapter's law?
20. DPDP Rule 6's security safeguards include:
21. The governance stack's deepest lesson:
22. LoanKart's CERT-In deadline, from 4:20 p.m. Saturday discovery:
23. LoanKart's DPDP Rule 7 duties in November 2026:
24. The audit verdict on LoanKart's bundled sign-up consent:
25. The vendor-owned model fails which principle?
26. The checklist ends with board reporting because:
27. DPDP erasure duties include warning the Data Principal:
28. Chapter 16's lending scandals meet this chapter at:
29. Chapter 32 takes up:
30. Why regulators set clocks as short as six hours: