CreditNirvana 安全事件復盤:兩項 API 漏洞對符合 DPDP 的 AI 部署有何啟示
CreditNirvana Breach Post-Mortem: What Two API Flaws Taught Us About DPDP-Compliant AI Deployment
CreditNirvana 兩項 API 漏洞令 6.6 million 個借款人帳戶及約 315,000 段催收通話錄音在沒有有效憑證下可被存取;兩項漏洞均於 2026 年 6 月 2 日修復。
CreditNirvana Breach Post-Mortem: What Two API Flaws Taught Us About DPDP-Compliant AI Deployment
In 2026 an independent security researcher found two flaws in CreditNirvana, an India-based agentic debt-collection platform owned by the fintech company Perfios. Together they made roughly 6.6 million borrower accounts and about 315,000 recorded collection calls reachable without valid credentials.
The vendor fixed both flaws on 2 June 2026, the same day CERT-In confirmed it would act. That response is worth stating plainly, because it is not the usual outcome. What follows is an attempt to separate what the research established from what it did not, and to draw out what it means for Indian firms now running AI agents over personal data.
The 90-Second Version
| Item | Detail |
|---|---|
| Platform | CreditNirvana (a Perfios company) — AI collections system for banks, NBFCs and asset reconstruction companies |
| Flaw 1 | CVE-2026-110474 — hardcoded API key in client-side code, Agentic AI Call Center (CVSS 8.7, High) |
| Flaw 2 | CVE-2026-110475 — improper authentication, Collections Platform (CVSS 9.3, Critical) |
| Scale | 6.6 million accounts across three clients; ~315,000 call recordings |
| Reported | 30 May 2026 via CERT-In; patched 2 June 2026 |
| Publicly disclosed | CVE note 26 October 2026; research published 18 November 2026 |
Two Vulnerabilities, Two Failure Classes
The disclosure describes two separate systems inside the same vendor, each defeated a different way. That distinction matters for assurance work, because the two failures are detectable by entirely different procedures.
Flaw 1 — A credential shipped to the browser
The Agentic AI Call Center authenticated users through Amazon Cognito. Two problems sat stacked on top of each other. An API key was hardcoded into client-side JavaScript, so it was delivered to every browser that opened the page. Separately, the Cognito token returned by the login flow was never used in subsequent requests — so the login step could be skipped by editing client code to supply plausible administrative values.
Per the disclosure, that access allowed a visitor to:
- listen to and read approximately 315,000 recorded collection calls, each analysed and transcribed
- review AI-generated post-call action recommendations
- browse and manage the platform's AI agent registry
- build and run their own AI-powered collection campaigns through the vendor's tooling
Flaw 2 — The token that was sent and never checked
The primary collections platform used Keycloak for identity management. This route is the more instructive of the two, because no secret leaked at all: the credential was present and simply not validated. It carried the higher CVSS score of the pair.
What Was Accessible
From the Aditya Birla Capital Finance instance alone, the researcher reported an inventory that makes the compliance analysis concrete:
| Data category | Volume observed |
|---|---|
| Borrower and business accounts | 3.4M (Aditya Birla Capital Finance) + 3.2M across two other instances = 6.6M |
| Settlement files including ID cards and full credit-bureau reports (one sample ran 26 pages) | More than 161,000 |
| Payment receipts, often with cheque images | ~3.7M (May 2026), ~3.6M (April 2026) |
| Voice-bot calls with recordings and AI summaries | 214,000 in May 2026 alone |
| Legal notices and lawyer case dashboards | 119,000 notices in May 2026 |
| SMS, email and messaging records | millions per month |
| Platform user accounts / collection agencies | 10,000+ users; 433 agencies |
Two rows go beyond data protection. Settlement files containing identity documents and credit histories are the raw material of identity fraud. Visit tracking for field collection agents is employee-surveillance data. Both belong in a different risk register from a marketing database.
Why "Agentic" Changes the Blast Radius
A conventional database leak is a read failure: the attacker obtains records. Here the accessible surface included instruments of action — campaign execution, agent management and legal escalation sat behind the same broken check as the data. When the system's selling point is that it decides "the next best action" and executes it autonomously, an authentication bypass is not only a disclosure risk. It is a control risk over the actions the agent can take on a borrower's account.
The DPDP Angle
Under India's Digital Personal Data Protection Act, obligations attach to the Data Fiduciary — the entity that determines why and how personal data is processed. In a collections scenario that is normally the lender, not the software vendor. The vendor's remediation speed does not discharge the lender's notification duty.
The DPDP Rules were notified on 13 November 2026, with obligations phased in through to 13 May 2027. The operational figure most analyses converge on is 72 hours from detection for notifying the Data Protection Board and affected Data Principals. For the affected lenders, the fix arriving quickly is good news about the vulnerability — it says nothing about whether their own filing obligations were met.
The CERT-In Clock Runs Faster
CERT-In's Directions from 2022 remain the tighter constraint: report specified cyber incidents within six hours of noticing them, and retain logs for 180 days. A firm that discovers exposure like this and cannot produce 180 days of relevant logs has a second problem on top of the first.
One timing detail a CA reviewing vendor contracts will recognise: five months passed between the fix (2 June) and the CVE note (26 October). Remediation speed and disclosure speed are separate measurements.
Confirmed, Unproven, and Unknown
| Status | Finding |
|---|---|
| Confirmed | Two vulnerabilities existed; CERT-In issued note CIVN-2026-0523 with CVE-2026-110474 and CVE-2026-110475; both were fixed on 2 June 2026; data was accessible without valid credentials during the window; exposure spanned 6.6 million accounts across three clients |
| Unproven | That any third party accessed, copied or misused the data. The published research does not claim this, and no affected institution has reported it |
| Unknown | How long the flaws were live (described as potentially years, not established precisely); whether the affected lenders filed CERT-In and DPDP notifications; the vendor's own account — CreditNirvana and Perfios did not respond to WIRED's requests for comment |
Five Things This Incident Demonstrates
- A secret in client-side code defeats authentication by definition. If the key is in the browser, the boundary is decorative.
- Sending a token is not validating a token. The more instructive flaw involved no leaked secret at all — the credential was present and ignored.
- Agentic systems make action, not just reading, the risk.
- One vendor's defect becomes several fiduciaries' obligation. The DPDP duties attach to the lender; the six-hour CERT-In clock starts on notice.
- Remediation speed and disclosure speed are separate metrics. A three-day fix and a five-month path to the CVE note can coexist.
What this article is not: it is not legal advice, and it does not assert that any party acted unlawfully, that data was exfiltrated, or that any affected institution failed its obligations. Where sources conflict or are silent, that is stated as such.
Sources
- Eaton Zveare, "The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms" — primary technical disclosure
- WIRED — AI debt collection service exposed 6.6 million accounts and 315,000 call recordings
- AccountsRecovery.net
- ReceivablesInfo
- CVE-2026-110474
- CVE-2026-110475
- CERT-In Directions summary (CyberSigma)
- SIRI Law LLP — India's CERT-In 6-hour incident reporting mandate
- Legismith Partners LLP — DPDP Act penalties explained
- CodeColonies — DPDP Act penalties explained
- Ampcus Cyber — RBI cybersecurity framework guide
來源:DEV Community · dev.to
