AI SecurityBreachesApplication Security

An AI Debt Collector Exposed 6.6 Million Accounts With Two Very Old Mistakes

November 21, 2026 · Brian Deitch

The pitch for CreditNirvana is agentic AI for debt collection: voice bots, co-browsing, automated recoveries. The security problem, disclosed this week by independent researcher Eaton Zveare, was not AI at all. It was a hardcoded API key and a backend that believed whatever username you sent it.

What Was Exposed

Per the researcher's disclosure as reported by NeoTeo and ExplainX:

  • About 6.6 million borrower accounts across three lenders and collectors in India: Aditya Birla Capital Finance (more than 3.4 million), and BharatPe and International Asset Reconstruction Company (about 3.2 million combined).
  • About 315,000 call recordings in a separate call-center system, along with transcripts.
  • Data types included ID cards, credit reports, legal and settlement documents, payment receipts, emails, text messages, call logs, and collection agents' location tracking.

How

Two separate flaws, tracked as CVE-2026-110474 and CVE-2026-110475:

  • Call-center system: an API key sitting in client-side JavaScript. Anyone who opened the browser's developer tools had access to recordings, transcripts and campaign functions.
  • Collections platform: authentication trusted a username in the request body instead of validating the login token. Change the name, become someone else. The reports say this also exposed admin access to the contact-center service behind it.

The timeline is the good news. The researcher reported it to CERT-In on May 30, CERT-In engaged within days, and both flaws were fixed on June 2. Credit to everyone involved for a fast fix. What we don't know is whether anyone else found it first; none of the coverage includes a statement from the company on that.

Why This Matters for AI Apps

I work at Zscaler, and I spend a lot of time talking to customers about AI risk. This story is a useful reality check. The scary AI headlines are about models doing clever things. The actual exposure here came from the same two bugs that have been on the OWASP list for 20 years.

AI platforms make the blast radius bigger, though. An agentic collections tool is wired into everything: credit bureaus, phone systems, messaging, document stores, agent location. That's what makes it useful, and it's why one broken auth check exposed call audio, credit reports and ID cards all at once.

What To Do

If you build or buy AI agent platforms:

  • No secrets in the browser. If the frontend can see a key, so can everyone. Put a backend in between and scope every token.
  • Identity comes from the validated token, never the request body. Test this explicitly: log in as user A, send user B's name, and see what happens.
  • Ask your AI vendors for their last pentest and specifically about broken object-level authorization and exposed keys. "We use AI" is not a security answer.
  • Inventory what your agents can reach. If one service account can read recordings, credit reports and IDs, that's the blast radius when it breaks.

If you're a borrower with these lenders:

  • Expect convincing collection scams. Someone with your loan details and call history can sound very legitimate. Verify any demand through the lender's official app or number before paying anything.

Bottom Line

Fixed fast, disclosed responsibly, and caused by bugs a basic API test would catch. The lesson isn't "AI is dangerous." It's that AI platforms still need the boring AppSec basics, and they connect to so much data that skipping them hurts more.

Sources: NeoTeo, ExplainX, TechBuzz

creditnirvanaperfiosagentic-aiapi-securityhardcoded-keysbroken-authenticationcve-2026-110474cve-2026-110475indiadata-exposure
All postsDiscuss on LinkedIn