The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard

CreditNirvana Flaws Exposed 6.6M Accounts, 315,000 Calls

By The Daily Byte ·20 Nov 2026 ·3 min read
Call centre worker wearing a headset at a computer in an office

Two security flaws in CreditNirvana, an Indian debt-collection platform that uses AI agents to chase overdue loans, could have given an attacker access to 6.6 million borrower accounts and about 315,000 recorded phone calls, according to an independent researcher. Both flaws have now been fixed, and there is no evidence that anyone other than the researcher used them.

Two flaws, two systems

Researcher Eaton Zveare published his findings on 18 November, in a write-up also covered by WIRED the same day. The first flaw sat in CreditNirvana's Agentic AI Call Center, a system used to send AI agents after delinquent loan customers. A secret API key was hardcoded in the client-side JavaScript of the web app, and the Amazon Cognito login token returned at sign-in was not used for the data calls at all. Anyone who read the page code could find the key and gain full access to the call centre: listening to and reading 315,000 recorded calls, viewing transcriptions and AI summaries, managing AI agents, and building call campaigns. The customers affected were borrowers of Aditya Birla Capital Finance.

The second flaw was in the main collections platform. It uses Keycloak for identity, but its own login call trusted the username sent in the request body and did not check the identity token sent alongside it. Combined with an unauthenticated endpoint that listed more than 10,000 users, that let Zveare sign in as a high-level user and take over the platform. India's cyber agency, CERT-In, assigned CVE-2026-110474 (8.7, High) to the call-centre flaw and CVE-2026-110475 (9.3, Critical) to the platform flaw, under vulnerability note CIVN-2026-0523.

What was within reach

The accounts belonged to customers of three lenders: Aditya Birla Capital Finance, with more than 3.4 million accounts, plus BharatPe and the International Asset Reconstruction Company, with a further 3.2 million between them. Inside an account, Zveare says he could view ID cards, full TransUnion CIBIL credit reports, legal notices and demand letters, payment receipts, emails and text messages, and call recordings with AI summaries. The platform also exposed the real-time location tracking of field debt-collection agents, and a feature to request legal action on an account. He also found admin access to the SlashRTC contact-centre system and a dashboard showing millions of messages sent to borrowers in a single month.

Reported in May, fixed the same week

Zveare reported the flaws through CERT-In on 30 May, and sent a proof-of-concept video on 1 June. CERT-In confirmed it would act on 2 June, and both vulnerabilities were fixed later that day. CERT-In issued its public vulnerability note on 26 October, and Zveare published his write-up on 18 November. WIRED said CreditNirvana and its owner, Perfios, did not answer its requests for comment. The reports do not say whether anyone other than Zveare used the flaws, and Zveare believes they may have been present for years, though no start date is given.

A warning for AI handling sensitive data

The case is a reminder that AI systems holding financial and personal data need the same basic login checks as any other system. Both flaws were simple oversights: a secret key left in public page code, and a login that trusted a username instead of checking the token. The encouraging part of the story is the response. Perfios fixed both flaws on the same day it learned of them, which is the part of the story other vendors should copy.