{"slug": "ai-debt-collection-platform-leak-exposes-6-6m-accounts", "title": "AI Debt-Collection Platform Leak Exposes 6.6M Accounts", "summary": "Security researcher Eaton Zveare disclosed two critical vulnerabilities in Credit Nirvana, the AI-powered debt-collection platform owned by Indian fintech Perfios, exposing roughly 6.6 million borrower accounts and about 315,000 recorded collection calls. An unauthenticated REST endpoint allowed sequential-ID enumeration of full borrower profiles, while a misconfigured S3-compatible bucket set to public-read exposed call archives under a predictable naming pattern. India's CERT-In classified both issues as critical, and the exposed data falls under India's Personal Data Protection Bill as sensitive personal data.", "body_md": "In early November 2026, independent security researcher Eaton Zveare disclosed two high‑severity vulnerabilities in Credit Nirvana, the AI‑powered debt‑collection platform owned by Indian fintech firm Perfios. The flaws granted unauthenticated access to **6.6 million borrower accounts** and approximately **315,000 recorded phone calls** from the platform’s agentic call centre. India’s Computer Emergency Response Team (CERT‑In) classified both issues as critical, prompting immediate public scrutiny and a rapid response from Perfios.\n\nCredit Nirvana markets itself as an autonomous decision‑engine that “reads a borrower’s full context, decides the next best action, and executes it autonomously, continuously, and within policy.” The platform is integrated into the collections pipelines of several of India’s largest fintech companies, handling millions of interactions each month. The breach therefore represents not only a massive data exposure but also a potential erosion of trust in AI‑driven financial services.\n\n🔹 **----------**\n\n  • Affected Component: -------------------\n\n  • Data Exposed: --------------\n\n  • Potential Exploitation: ------------------------\n\n🔹 **Critical**\n\n  • Affected Component: REST API endpoint for borrower profile retrieval\n\n  • Data Exposed: Personal identifiers, loan amounts, repayment history\n\n  • Potential Exploitation: Unauthenticated GET requests could enumerate accounts and harvest PII\n\n🔹 **Critical**\n\n  • Affected Component: Internal call‑recording storage service (S3‑compatible bucket)\n\n  • Data Exposed: 315 k audio recordings, timestamps, agent IDs\n\n  • Potential Exploitation: Misconfigured bucket permissions allowed public read access\n\nThe first vulnerability stemmed from an improperly scoped authentication middleware. The `/v1\n\n/v1/borrower/{id}` endpoint was left without any token verification, allowing anyone who could guess or enumerate numeric IDs to retrieve full borrower profiles. The route accepted a simple integer identifier and returned a JSON payload containing the borrower’s name, PAN number, Aadhaar‑linked phone, email address, loan amount, outstanding balance, repayment schedule, and even the internal risk score used by the AI engine to prioritize collection actions. Because the IDs were sequential and started at 1, an attacker could script a rapid scan that harvested millions of records in under an hour.\n\nThe flaw was compounded by a secondary issue in the pagination logic: the API ignored the `limit` parameter and defaulted to a maximum of 10,000 records per request, but the `offset` parameter could be set to any integer. By iterating the offset in steps of 10,000, the researcher demonstrated how the entire dataset could be exfiltrated with just a few hundred HTTP calls.\n\nThe second vulnerability involved the storage layer for voice recordings. Credit Nirvana archives every outbound and inbound collection call in an S3‑compatible object store, tagging each file with metadata such as `call_id`, `borrower_id`, `agent_id`, and timestamps. The bucket’s ACL (Access Control List) was mistakenly set to `public-read`, meaning any URL pointing to an object could be accessed without authentication.\n\nThe researcher discovered that the bucket’s naming convention followed a predictable pattern: `cn-call-recordings-prod-<region>-<date>.zip`. By enumerating dates and regions, they were able to download entire daily archives containing thousands of recordings. Within the recordings, borrowers could be heard discussing sensitive financial details, personal circumstances, and even health information—data that falls under India’s Personal Data Protection Bill (PDPB) as “sensitive personal data”.\n\n🔹 **--------**\n\n  • Estimate: ----------\n\n🔹 **Borrower accounts exposed**\n\n  • Estimate: 6,600,000\n\n🔹 **Call recordings exposed**\n\n  • Estimate: 315,000\n\n🔹 **Average PII fields per account**\n\n  • Estimate: 12\n\n🔹 **Potential regulatory fines (PDPB)**\n\n  • Estimate: Up to ₹5 crore per breach\n\n🔹 **Estimated remediation cost**\n\n  • Estimate: ₹12–15 crore\n\nThe scale of the exposure is staggering for a single fintech service. Beyond the immediate privacy concerns, the breach raises questions about the robustness of AI‑driven decision pipelines that rely on accurate, up‑to‑date data. If malicious actors were to manipulate the exposed data—e.g., by injecting false repayment histories—they could potentially influence the AI’s risk scoring, leading to wrongful collection actions or denial of credit.\n\nPerfios issued a public statement within 24 hours of the CERT‑In advisory:\n\n“We take the security of our customers’ data with the utmost seriousness. Upon learning of the vulnerabilities, we immediately disabled the affected API endpoints, rotated all authentication keys, and applied a restrictive bucket policy to the call‑recording storage. We have also engaged a third‑party security firm to conduct a full forensic audit and will notify affected borrowers in accordance with applicable law.”\n\nThe company also pledged to:\n\nCERT‑In classified both findings as **Critical (CVSS 9.8)** and recommended that any organization using Credit Nirvana perform an immediate security review.\n\nThe Credit Nirvana incident underscores a recurring theme in the fintech sector: the race to ship AI‑powered products often outpaces the implementation of foundational security controls. While the platform’s autonomous decision‑making capabilities promise efficiency gains, they also create a larger attack surface:\n\n*Read the full breakdown originally published at [https://ltdeveloperblogs.github.io/posts/an-ai-debt-collection-service-exposed-66-million-accounts-and-315000-call-recordings/](https://ltdeveloperblogs.github.io/posts/an-ai-debt-collection-service-exposed-66-million-accounts-and-315000-call-recordings/)*", "url": "https://wpnews.pro/news/ai-debt-collection-platform-leak-exposes-6-6m-accounts", "canonical_source": "https://dev.to/luckytaorem/ai-debt-collection-platform-leak-exposes-66m-accounts-321i", "published_at": "2026-11-22 12:09:32+00:00", "updated_at": "2026-11-22 12:12:23.286548+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "ai-agents", "ai-ethics"], "entities": ["Credit Nirvana", "Perfios", "Eaton Zveare", "CERT-In", "India"], "also_reported_by": [], "takeaway": "An unauthenticated borrower-profile endpoint with sequential integer IDs and an offset-iterable pagination flaw let a researcher harvest 6.6 million records in a few hundred HTTP calls, while a public-read S3 bucket with predictable date-and-region naming exposed 315,000 call recordings.", "recommendation": {"advice": "Enforce token verification on every API route, use non-sequential identifiers with capped pagination, and audit object-store ACLs and bucket naming so recordings are never publicly readable.", "playbook": "agent-security", "playbook_url": "https://wpnews.pro/playbooks/agent-security"}, "alternates": {"html": "https://wpnews.pro/news/ai-debt-collection-platform-leak-exposes-6-6m-accounts", "markdown": "https://wpnews.pro/news/ai-debt-collection-platform-leak-exposes-6-6m-accounts.md", "text": "https://wpnews.pro/news/ai-debt-collection-platform-leak-exposes-6-6m-accounts.txt", "jsonld": "https://wpnews.pro/news/ai-debt-collection-platform-leak-exposes-6-6m-accounts.jsonld"}}