Independent security researcher Eaton Zveare revealed two security bugs in CreditNirvana that exposed 6.6 million user accounts and 315,000 call recordings on Wednesday, November 18. The CreditNirvana data exposure left personal files, audio archives, and recovery suggestions open on the web without any login check [1]. Owned by Indian fintech firm Perfios, CreditNirvana uses AI tools to manage debt recovery for major Indian banks. The leak shows how basic coding bugs can break customer data systems when lenders hand private records to third-party tech vendors [2].
CreditNirvana Data Exposure Across Three Lenders
CreditNirvana runs an AI debt platform built to scan loan records, pick collection steps, and reach debtors through voice bots and text alerts. Built by parent company Perfios in India, the AI platform aims to “read a borrower’s full context, decide the next best action, and execute it autonomously, continuously, and within policy” [1]. But security flaws found by Eaton Zveare left millions of private loan files open on the web for three major Indian client lenders [2]. Reporter Lily Hay Newman at WIRED noted that the Indian firm built its AI tools to automate debt collection at scale, yet simple flaws left borrower files exposed [1]. The leak showed how much private data banks entrust to outside tech firms.
Three large finance firms accounted for all 6.6 million exposed accounts in the leaky database. Aditya Birla Capital Finance alone accounted for over 3.4 million debtor accounts and all 315,000 call recordings stored in the system. Indian fintech firm BharatPe and International Asset Reconstruction Company made up the remaining 3.2 million exposed records [2]. Zveare told WIRED that big Indian lenders lean heavily on the vendor: “Some of the largest fintech firms across India use this company to autonomously collect debts, so they naturally hold data, including credit scores, call recordings, transcripts, court records, and other sensitive documents. There’s just so much in there that makes it a very sensitive issue” [1].
Debtors had no direct contracts with CreditNirvana, yet their loan files moved to its servers through bank outsourcing [1]. Handing those private files to Perfios created new risks that borrowers could not see or control [2]. A single bad endpoint opened millions of files to any browser.
Hardcoded Keys and Bypassed Token Checks
The flaws in CreditNirvana did not rely on custom exploit tools. Writing for Startup Fortune, Julian Lim said that anyone opening developer tools in a web browser could find the first bug in minutes. In the call center system, an API key sat hardcoded in client JavaScript. By tweaking the page in a browser, Zveare bypassed Amazon Cognito login checks to view call audio without a password [2].
The second bug, located on the main CreditNirvana platform, was worse because the server never checked Keycloak login tokens. Instead of verifying digital signatures, the server accepted any username typed into the request body. An intruder could submit an admin username and gain full administrative rights [2]. Cybersecurity team India Computer Emergency Response Team CERT-In classified this flaw as critical severity, warning that an attacker could impersonate administrators and alter files. CERT-In noted that intruders could also track the real-time physical locations of human debt agents working in the field across Indian cities [1].

In its formal advisory, CERT-In explained the core flaw: “Successful exploitation of this vulnerability could allow the attacker to enumerate platform users, establish sessions as arbitrary users including administrative users, access and modify sensitive information with the privileges of the impersonated user” [1]. The bug left core business controls open to the web.
Borrower Files Stored on Leaky Systems
The exposed files went far beyond simple name lists. The open call center held about 315,000 recorded calls between loan agents and Aditya Birla Capital Finance debtors. Each call file included speech transcripts, sentiment scores rating debtor distress, and automated next-step advice [2]. Anyone finding the link could listen to people facing money troubles.
On the main CreditNirvana platform, the exposed storage held deep loan dossiers used for debt recovery. Files open through broken Keycloak token checks included government identity cards, credit bureau files, court documents, case notes, and legal records. The system also stored millions of written notices, including email logs and SMS messages sent to debtors about unpaid balances [1]. Because Perfios pulled diverse loan records into one searchable index for Aditya Birla Capital Finance, BharatPe, and International Asset Reconstruction Company, a single flaw opened all those files at once [2].
A big danger involved taking control of the AI debt bots in the call center dashboard. Zveare found that an attacker could change bot prompts and run rogue call campaigns [1]. As Julian Lim wrote in Startup Fortune, taking control of outbound voice bots could lead to direct fraud: “Controlling the agent that phones debtors is worse, because it puts an attacker’s script in a collector’s voice, talking to people already primed to pay whoever sounds official” [2].

CreditNirvana Vulnerabilities Found and Fixed
Fixing the flaws moved quickly after Eaton Zveare sent his report to India Computer Emergency Response Team CERT-In on May 30, 2026. CreditNirvana rolled out software patches on June 2, closing both bugs three days after CERT-In warned Perfios leaders [2]. Speaking to WIRED, Zveare praised the quick patch: “They fixed it in one day, which was great, but this is really a pretty significant case”. He warned that the bugs had likely sat open in CreditNirvana code for years before anyone checked [1].
Formal tracking records arrived months later from cybersecurity bodies. On October 26, 2026, the bugs received official tags as CVE-2026-110474 and CVE-2026-110475. Public disclosure followed on November 18, ending a 172-day coordination period between Zveare and CERT-In [2]. Neither Perfios nor CreditNirvana answered questions from WIRED about data security practices [1].
It remains unknown whether rogue actors accessed the data before June. Neither Perfios nor Aditya Birla Capital Finance said whether they ran forensic checks for past leaks. Furthermore, none of the three client lenders confirmed sending alert notices to debtors whose files sat exposed online [2].
Lender Accountability in AI Outsourcing
The CreditNirvana data exposure raised new questions for banks hiring AI startups for core operations. While lenders can hand daily call workflows to AI platforms, they can’t hand off legal blame to debtors or regulators when security fails. Lenders face rising regulatory scrutiny over automated systems, as seen when US banks restrict unauthorized AI agents over account risks to protect user balances. When tech vendors make basic login errors, client banks face the fallout [2].
Similar data exposures have hit other firms using automated support bots. Earlier in 2026, Sears Home Services exposed nearly 3.7 million customer records through open AI bot databases, as SC Media reported. In another breach reported by Cybernews, an open cloud bucket stored 55,000 audio files of staff speaking with student debt holders, leaking Social Security numbers in the recordings and forms [2]. In each case, AI tools grouped scattered files into one place, multiplying the damage when server checks failed.
As AI platforms spread in debt recovery, banks must verify where access checks occur and set clear limits on how long voice recordings stay on file [2]. Zveare said the CreditNirvana data exposure serves as “an important example of an AI-powered service having a very sensitive data exposure” [1]. Until banks test third-party AI software with rigor, debtor records will remain at risk.
- ONLINE NEWS Newman, L. H. (2026, November 18). An AI debt collection service exposed 6.6 million accounts and 315,000 call recordings. WIRED. [Article Link]
- ONLINE NEWS Lim, J. (2026, November 19). Two flaws in CreditNirvana’s AI collections platform exposed 6.6 million accounts. Startup Fortune. [Article Link]