Confirmed: two vulnerabilities in CreditNirvana, an India-based AI debt-collection platform owned by the fintech company Perfios, made 6.6 million borrower accounts and roughly 315,000 collection-call recordings reachable without authentication. Independent security researcher Eaton Zveare reported them through CERT-In on 30 May 2026, and both were fixed on 2 June 2026 — the same day CERT-In confirmed it would act.
Not established: that anyone other than the researcher accessed that data. Neither Zveare's write-up nor WIRED's reporting indicates any third party used the flaws, and the researcher's own summary states the investigation proved the information was accessible, not that it was extracted. CreditNirvana and Perfios did not respond to WIRED's requests for comment.
For chartered accountants, the interesting part is not the breach headline. It is the shape of the two failures, both of which sit squarely inside the assurance questions a CA firm now gets asked when a client deploys an AI agent against customer data — and both of which map onto obligations that already exist in Indian law.
💡 Why It Matters for Indian CAs
Debt-collection platforms sit on the most sensitive document set in Indian retail finance: ID cards, full credit-bureau reports, settlement files with cheque images, and call recordings. When a platform serves multiple regulated lenders at once, one authorization defect becomes a multi-client exposure. The DPDP Act's security-safeguard and breach-notification duties attach to the Data Fiduciary — the client, not the vendor — which is why vendor assurance questions now carry statutory weight.
The 90-Second Version
| Item |
Detail |
| Platform |
CreditNirvana (a Perfios company) — AI collections platform for banks, NBFCs and asset reconstruction companies |
| Flaw 1 |
CVE-2026-110474 — hardcoded API key in client-side code, Agentic AI Call Center (8.7, High) |
| Flaw 2 |
CVE-2026-110475 — improper authentication, Collections Platform (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 |
| Disclosed publicly |
CVE note 26 October 2026; research published 18 November 2026 |
What Actually Happened: Two Vulnerabilities, Two Failure Classes
Zveare's write-up 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 completely different procedures.
Flaw 1 — A Credential Shipped to the Browser
The Agentic AI Call Center authenticated users through Amazon Cognito. The researcher found two problems stacked on top of each other. First, an API key was hardcoded into client-side JavaScript — meaning it was delivered to every browser that opened the page. Second, the Cognito token returned by the login flow was not used anywhere in subsequent requests. Because the returned credential was never checked, the login step could be skipped entirely by editing the client code to supply plausible administrative values.
What that access allowed, per the disclosure:
- Listening to and reading approximately 315,000 recorded collection calls, each analysed and transcribed.
- Reviewing AI-generated post-call action recommendations — the platform's "next best action" output for each borrower.
- Browsing and managing the platform's AI agent registry.
- Building and running their own AI-powered collection campaigns through the vendor's own tooling.
The affected client on this system was Aditya Birla Capital Finance, whose customers' recordings and transcripts were in scope.
Flaw 2 — The Token That Was Sent and Never Checked
The primary collections platform — the instance for Aditya Birla Capital Finance is named CollectionWorkz — used Keycloak for identity management. The researcher's route through it is the more instructive of the two, because it is a logic defect rather than a leaked secret.
An API call to a users-list endpoint returned more than 10,000 platform users without authentication. From there, the researcher found that the application transmitted the Keycloak token in an Authorization-Token header but never validated it. The platform instead trusted a userName value supplied in the request body. Substituting a username with a high-level role produced a successful administrative login. A second token (CN-Token) was required on later requests, but once the forged headers were removed, those requests succeeded too.
⚠️ The Transferable Lesson
A hardcoded key is a secret-management failure. A token that is transmitted but never verified is an authorization-design failure. The second class is invisible to secret scanners, because there is no secret to find — the credential is present and simply ignored. Any review that stops at "are there hardcoded secrets?" would have cleared this platform.
The Disclosure Timeline
The remediation record here is unusually clean, and it is worth stating plainly because it is the part the vendor got right.
| Date (2026) |
Event |
| 30 May |
Vulnerabilities reported to CERT-In |
| 1 June |
CERT-In requests a proof-of-concept video; researcher supplies it |
| 2 June |
CERT-In confirms action; both vulnerabilities fixed the same day |
| 26 October |
CERT-In vulnerability note CIVN-2026-0523 issued; CVE-2026-110474 and CVE-2026-110475 assigned |
| 18 November |
Technical write-up published; WIRED reports the findings |
Zveare's account credits the vendor: the fixes landed the same day the company learned of the issues. He also told WIRED the flaws were potentially present for years, and that the case remains significant because so much sensitive material sat behind a single authorization check. CERT-In's advisory recommends upgrading the call centre to version 2.0.0 and the collections platform to version 4.8.2.
One timing observation that 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 different measurements.
What Was Accessible
The exposure inventory is the part that makes the DPDP analysis concrete, because "personal data" is easy to discuss in the abstract and hard to argue with when itemised. From the Aditya Birla Capital Finance instance alone, the researcher reported:
| Data category |
Volume observed |
| Borrower and business accounts |
3.4 million (Aditya Birla Capital Finance) + 3.2 million across two other instances = 6.6 million |
| Settlement files, including ID cards and full credit-bureau reports (one sample report ran 26 pages) |
More than 161,000 |
| Payment receipts, frequently with cheque images |
~3.7 million (May 2026) and ~3.6 million (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 of records per month |
| Platform user accounts; collection agencies on the platform |
More than 10,000 users; 433 agencies |
| Field-collector visit tracking (who was visited, and the outcome) |
Present and enumerable in the interface |
Two entries in that table go beyond data protection. Settlement files with identity documents and credit histories are the raw material of identity fraud. Location tracking of field collection agents is employee-surveillance data about the vendor's own or its clients' staff. Both sit in a different risk register from a marketing database.
Why "Agentic" Changes the Blast Radius
The CreditNirvana platform is described, in its own words, as reading "a borrower's full context", deciding "the next best action", and executing it "autonomously, continuously, and within policy." That is the marketing description of agentic design — and it is exactly what makes the failure mode different from a conventional data breach.
A classic database leak is a read failure. The attacker obtains records. Here, the accessible surface included instruments of action:
Read exposure
Call recordings, transcripts, AI post-call recommendations, credit reports, identity documents, receipts, legal notices, messaging history, user lists.
Action exposure
The AI agent registry (browse and manage agents), campaign construction and execution, receipt approval or rejection, and a function to request legal action on an account.
The second column is the new part. An actor with that access does not merely learn about a borrower — the actor can direct automated systems to contact that borrower, and can move an account toward legal escalation. Whether anyone exercised those capabilities is unknown and unproven; what matters structurally is that the capability existed and was reachable without a valid credential.
This is the same class of finding PyBankers reported in the PixelLeak disclosure, where AI coding agents leaked more than 13,000 screenshots: the agent is granted broad access to be useful, and the blast radius is set by that grant, not by the data model.
The DPDP Angle: What the Act Requires of a Data Fiduciary
India's Digital Personal Data Protection Act, 2023 supplies two obligations that this incident maps onto directly. Both are in section 8, and both carry penalties set out in the Schedule to the Act.
| Obligation |
Provision |
Maximum penalty |
| Reasonable security safeguards to prevent a personal data breach |
Section 8(5) |
₹250 crore |
| Notifying the Data Protection Board and affected Data Principals of a breach |
Section 8(6) |
₹200 crore |
Legal analyses of the Act consistently note that these penalties attach per violation and can apply together, and that the Board weighs factors including the nature and severity of the breach, the data involved, whether it is a repeat occurrence, and the mitigation the fiduciary undertook. Timely notification and documented safeguards are repeatedly described as the mitigators a fiduciary controls.
The Rules notified on 13 November 2025 put the framework on a phased clock. Analyses of the phase dates place consent-manager integration obligations at around 13 November 2026 and the substantive fiduciary duties — including security safeguards and breach notification — at around 13 May 2027, subject to amendment. Practically, that means the security-safeguard standard is being tested in public incidents now while its full compliance date is still ahead.
On timing, the operational figure most analyses converge on is 72 hours from detection for notifying the Board and affected Data Principals. The obligation itself is on 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; it only affects how quickly the lender learns it has one.
The CERT-In Clock Runs Faster Than the DPDP Clock
CreditNirvana was reported through CERT-In, which is why the case contains a second, tighter compliance thread. CERT-In's Directions, issued on 28 April 2022 under section 70B(6) of the Information Technology Act, 2000 and effective from late June 2022, impose requirements that bite harder than the DPDP timeline:
- Six hours to report any of the incident categories listed in Annexure I to the Directions, counted from the moment the incident is noticed or brought to the entity's notice.
- 180 days of logs for all ICT systems, retained securely within Indian jurisdiction, and produced to CERT-In with an incident report or on demand.
- Clock synchronisation to NIC, NPL or traceable NTP sources.
Analyses of section 70B(7) note that non-compliance carries criminal liability on designated individuals — imprisonment of up to one year, or a fine of up to ₹1 lakh, or both — rather than only organisational consequences. For entities in the regulated financial sector, CERT-In reporting sits alongside parallel sector reporting: analyses of the RBI framework describe incident reporting to the RBI through DAKSH within six hours of detection where the applicable Direction requires it, with housing finance companies reporting to the NHB under the relevant provisions.
💡 The Two-Clock Sequence
CERT-In: six hours. DPDP: commonly described as 72 hours. Financial-sector regulators: a parallel six-hour window where applicable. These are separate filings to separate authorities, triggered by the same event, and none of them is satisfied by telling the other.
The Third-Party Question: Fiduciary, Processor, and the Contract Between Them
The structural fact of this case is that one vendor's authorization defect exposed data belonging to multiple financial institutions at once — three named clients across separate instances, with 3.4 million accounts in one and 3.2 million in the others. A shared collections platform concentrates risk in a way that a per-lender deployment does not.
Under the DPDP framework, an entity processing personal data on a fiduciary's behalf is a Data Processor, and the processor's obligations are set largely by contract with the fiduciary. The fiduciary's statutory duties do not migrate to the processor. This is the seam where assurance practice and data-protection law meet, and it produces questions with unusually concrete answers in a post-mortem:
- Notification chain: what is the contractual maximum time between the vendor's detection and the client's notification? Does it fit inside a six-hour CERT-In filing and a 72-hour DPDP filing?
- Authentication standard: does the contract specify server-side validation of tokens, or does it specify an outcome ("access control") that a client-side check can be claimed to satisfy?
- Secret handling: is there a defined prohibition on shipping credentials to client-side code, and a testing obligation that would detect it?
- Tenant isolation: for a multi-client platform, what prevents a defect in one instance from being reachable from another, and what evidence is produced?
- Logging and attribution: are API keys, service accounts and AI agent identities logged with attribution, and are those logs inside India for 180 days?
- Exit and deletion: what happens to recordings, transcripts and derived AI outputs on termination?
PyBankers examined the adjacent ground for CA firms in its review of 150 AI tools for chartered accountants, and the CERT-In advisory angle in the CRAFTY FALCON advisory analysis. The vendor-assurance questions above are the same questions in a procurement context.
An Agentic Data-Exposure Threat Model
The reusable output of this incident is a small set of failure classes. Each is stated below as a condition to test rather than a control to buy, because the appropriate response depends on the system and the reader's own risk appetite.
1. Credential placement
- Is any credential that grants access to personal data delivered to a client-side context?
- If a returned authentication token exists, is it enforced on every subsequent request — or is some other request field trusted instead?
2. Authorization as a server-side decision
- Does the server independently verify the identity claim it is given, or does it accept an identity asserted in the request body?
- Are there endpoints that return data before any authentication step — for example, directory or user-list calls?
- Do downstream endpoints re-validate, or do they rely on an earlier step having validated?
3. Agent permissions as the new perimeter
- What can the agent do, not just read: which outbound communications, which workflow state changes, which escalations?
- Is there a registry of agents, and is it protected by the same authorization as the data itself?
- Can campaign parameters be altered through the same interface that reports on campaigns?
4. Derived data and secondary artefacts
- Transcripts, sentiment labels and AI post-call recommendations are new stores that did not exist before the agent was deployed. Where do they live, who can read them, and what is their retention period?
- Do these derived stores inherit the access controls of the source system, or their own weaker ones?
5. Multi-tenant blast radius
- How many clients share the deployment, and is isolation architectural or configuration-based?
- Does a single authorization defect reach one tenant's data or every tenant's data?
What an Assurance Review Would Test
This is the section a CA firm is most likely to be asked for, so it is worth being precise about what kind of work this is. Neither the DPDP Act nor CERT-In's Directions prescribe an audit procedure. What follows are evidence requests that map to the failure classes above — the professional judgement about scope, materiality and opinion belongs to the engagement.
| Question |
Evidence that answers it |
| Are credentials exposed to the client? |
Static analysis of shipped JavaScript bundles; CI check that fails a build on embedded keys |
| Is the returned token actually enforced? |
Test case: present a valid token with a mismatched identity claim, and an invalid token with a valid identity claim. Record both outcomes |
| Do any endpoints return data pre-authentication? |
Unauthenticated enumeration of the API surface, with the resulting response codes retained as evidence |
| What can an agent be directed to do? |
A permissions matrix for agent identities across read, communicate, transact and escalate actions |
| Are derived AI artefacts governed? |
Data-flow inventory covering transcripts, summaries and recommendation outputs, with retention and access rules |
| Is logging sufficient for a six-hour filing? |
Evidence of 180-day retention within India, NTP synchronisation, and a timed drill from detection to a filed report |
| Does the contract support the fiduciary's duties? |
Clauses covering notification windows, security standards, multi-tenant isolation, audit rights and deletion on exit |
For the audit-adjacent version of this problem — where AI is applied to public-sector assurance rather than to customer collections — PyBankers covered the CAG's in-house LLM work at the Madras audit office. And where the objective is to avoid transmitting sensitive documents at all, the 13 MB offline GST extraction engine is a worked example of local-only processing.
What This Article Is Not
⚠️ Scope Limits
This is a factual post-mortem of a published vulnerability disclosure, not legal advice, an audit opinion, or a certification of any party. The technical findings are Eaton Zveare's, reproduced with attribution; the quantified exposure figures are his observations as reported. Nothing here establishes that any data was copied by anyone other than the researcher, and none of the affected institutions has been shown to have acted improperly. Statutory references to the DPDP Act, 2023 and the CERT-In Directions are cited for orientation; the phase dates of the DPDP Rules are described as they appear in published analyses and remain subject to amendment. Readers evaluating their own vendor arrangements should rely on the primary sources listed below and on their own advisers.
Confirmed, Unproven, and Unknown
Separating these three categories is the discipline that keeps a post-mortem useful rather than inflammatory.
| Status |
Finding |
| Confirmed |
Two vulnerabilities existed; CERT-In issued vulnerability note CIVN-2026-0523 with CVE-2026-110474 and CVE-2026-110475; both were fixed on 2 June 2026; the data was accessible without valid credentials during the window; the 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 of the events — CreditNirvana and Perfios did not respond to WIRED's requests for comment |
Key Takeaways
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 of the two flaws had no leaked secret at all — the credential was present and ignored.
- Agentic systems make action, not just reading, the risk. Campaign execution, agent management and legal escalation sat behind the same broken check as the data.
- 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 one-day fix and a five-month path to the CVE note can coexist.
Sources
Eaton Zveare, "The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms" — primary technical disclosure; CVE identifiers, remediation timeline, exposure inventory (checked 22 Nov 2026)
WIRED, "An AI Debt Collection Service Exposed 6.6 Million Accounts and 315,000 Call Recordings" — researcher quotes, platform description, non-response from CreditNirvana and Perfios (checked 22 Nov 2026)
AccountsRecovery.net — severity scores, per-instance account counts, statement that no third-party access is indicated (checked 22 Nov 2026)
ReceivablesInfo — CERT-In advisory references, recommended upgrade versions, and the accessibility-versus-exfiltration distinction (checked 22 Nov 2026)
CVE-2026-110474 — Agentic AI Call Center (8.7, High); CVE-2026-110475 — Collections Platform (9.3, Critical) (checked 22 Nov 2026)
CERT-In Directions summary (CyberSigma) — six-hour reporting window, 180-day log retention, NTP synchronisation, section 70B(6) basis (checked 22 Nov 2026)
SIRI Law LLP analysis — section 70B(7) penalties and the applicability of the Directions (checked 22 Nov 2026)
Legismith Partners LLP, "DPDP Act Penalties Explained" — penalty schedule with section 8(5) and 8(6) maxima and the factors the Board weighs (checked 22 Nov 2026)
CodeColonies analysis — DPDP Rules notification date of 13 November 2025 and the phased compliance dates (checked 22 Nov 2026)
Ampcus Cyber, RBI cybersecurity framework guide — six-hour incident reporting through DAKSH for covered regulated entities, alongside CERT-In obligations (checked 22 Nov 2026)
Related Reading from PyBankers
Last updated: 22 November 2026
📌 Editorial Disclosure & Verification
Researched and drafted with PyBankers AI research pipelines; fact-checked, edited, and verified by PyBankers Team. This report is for educational and informational purposes only and does not constitute financial, legal, or investment advice. Verify critical details independently.