cd /news/ai-agents/the-ai-repo-man-exploiting-agentic-a… · home › topics › ai-agents › article
[ARTICLE · art-172035] src=eaton-works.com ↗ pub= topic=ai-agents verified=true sentiment=↓ negative

The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms

Security researchers disclosed two vulnerabilities in Credit Nirvana's AI-powered debt collection platforms, tracked by India's CERT-In as CIVN-2026-0523: CVE-2026-110474 (rated 8.7 High) and CVE-2026-110475 (rated 9.3 Critical). A hardcoded API key in client-side JavaScript gave full access to the Agentic AI Call Center used by Aditya Birla Capital Finance, exposing 315,000 call recordings and the ability to build AI-agent call campaigns, while improper authentication let attackers bypass Keycloak and log in as admin to the primary platform, exposing 6.6 million accounts across Aditya Birla Capital Finance, BharatPe and International Asset Reconstruction Company (IARC) containing ID cards, TransUnion CIBIL credit reports, legal documents and agent location tracking.

by read10 min views2 publishedNov 18, 2026
The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms
Image: Eaton-Works (auto-discovered)

Key Points / Summary #

  • Credit Nirvana, an AI-powered debt collections platform in India, had 2 vulnerabilities:
    1. A hardcoded API key in client-side Javascript enabled full access to the Agentic AI Call Center.
    2. Improper authentication in the primary platform made it possible to log in as admin and gain access to all features.
  • Impact Highlights – Agentic AI Call Center:
    • Aditya Birla Capital Finance customers impacted.
    • Access to listen to/read 315k call recordings.
    • Review post-call recommendations to carry out actions.
    • Build and manage AI-powered campaigns.
    • Browse and manage AI agents.
  • Impact Highlights – Primary Platform:
    • Aditya Birla Capital Finance, BharatPe, International Asset Reconstruction Company (IARC) customers impacted.
    • 6.6 million accounts across those 3 clients, all containing sensitive information such as ID cards, credit reports, legal documents, and communication activity. There was also a feature to request legal action on an account.
    • Admin access to SlashRTC.
    • Millions of emails and text messages sent to customers.
    • Location tracking of debt collection agents.
    • Call recordings and AI summaries.
    • Legal notices, case management.
[CERT-In Vulnerability Note (CIVN-2026-0523)](https://www.cert-in.org.in/s2cMainServlet?pageid=PUBVLNOTES01&VLCODE=CIVN-2026-0523)

1. [CVE-2026-110474](https://www.cve.org/CVERecord?id=CVE-2026-110474) (8.7 / High)
2. [CVE-2026-110475](https://www.cve.org/CVERecord?id=CVE-2026-110475) (9.3 / Critical)

Before the advent of AI, if you missed a loan payment, you might get a friendly phone call reminder, or a collections agent demanding payment at your door. Enter 2026, agentic AI has invaded the debt/collections industry, and odds are the collections agent that calls you next may actually be an AI agent instead of a real person.

What could be worse than an agentic AI debt collection system? A hacked agentic AI debt collection system! In this vulnerability disclosure blog, we will dive into 2 systems that were vulnerable to complete takeover:

  1. Credit Nirvana’s Agentic AI Call Center – a system used by Aditya Birla Capital Finance to send AI agents after their delinquent loan customers.
  2. Credit Nirvana’s Primary Platform – a comprehensive platform used by India’s largest leading financial institutions to manage collections.

Credit Nirvana #

Credit Nirvana, a Perfios company, is a rapidly growing digital-first debt collection company in India. It is currently pioneering AI-powered debt collection systems with India’s leading financial institutions. This rapid advancement has come at a cost: the primary platform had a critical vulnerability that made it possible to bypass Keycloak and log in as admin to gain access to the sensitive information of millions. From ID cards, legal demand letters, full TransUnion CIBIL credit reports, real-time tracking systems, and even call recordings, troves of sensitive information were left exposed by an advanced platform with unfortunate security oversights. In addition, their Agentic AI Call Center had a hardcoded API key that could be used to access hundreds of thousands of call recordings for Aditya Birla Capital Finance customers and set up call campaigns powered by AI agents.

Credit Nirvana Agentic AI Call Center #

Did you roll your eyes as soon as you read that? Get ready to roll them again.

When you visit the site, you are asked to sign in. There is no registration function.

Looking into the code, it used Amazon Cognito to handle authentication. I came up with some modifications to skip the signin API call and trick it into believing a valid admin just signed in. The API call is commented out and the values that were expected in the response were replaced with plausible admin values:

With those modifications in place, you could log in using any email and password.

There’s some really interesting stuff here. Let’s take a look!

Note: Aditya Birla Capital Finance customers are the ones impacted.

Call Transcriber Agent

“Real-time call transcription with language detection and sentiment analysis”

There are 315k phone calls recorded here. They are analyzed, transcribed, and you can view them all and listen to them!

Disposition Agent

“AI-powered call disposition analysis and insights”

Post Call Action Recommender

“AI-powered next best action recommendations based on call analysis”

Call Audit Agent

This is where the “Agentic AI Calls” are audited. These shouldn’t be a thing, but they are.

Campaigns

“Build an AI-powered campaign to proactively engage customers before their payments become overdue”

“Monitor and manage your AI-powered campaigns in real-time”

AI Agent Registry

“Browse and manage AI agents available in the CreditNirvana collection system”

Hardcoded API Key

This was all made possible thanks to a hardcoded API key:

The Amazon Cognito token returned from the login was not used at all, which is why the login could simply be skipped.

Credit Nirvana Platform #

The Agentic AI Call Center was only the tip of the iceberg. The primary platform is much bigger and is jam packed with all sorts of data and features. I came across Aditya Birla Capital Finance’s Credit Nirvana instance which is called CollectionWorkz – their enterprise-wide digital platform for collections across secured and unsecured products.

When you visit the site, you are asked to sign in. There is no registration function.

This is a ReactJS app, so naturally I dug into the code to see what the APIs looked like, specifically any involving users.

By constructing an API call to that users-list endpoint, I was surprised to see it return a list of more than 10k users – all without being authenticated!

This was a promising start and if this API was vulnerable, there were likely many more.

Uncloaking Keycloak

The platform uses Keycloak for their IAM. My goal was to find a way to trick it into thinking a user was logged in, to see how far things could go. The first step was to stop it from redirecting you back to the login page. I found that setting some syntactically valid JWTs was enough to make it think a valid session might exist:

Predictably, it tried using the refresh token to get a new access token – and it fails:

To stop this, I killed the function that handles the refresh:

Then it tried to obtain some user info using the token, but that also fails since the token is no good:

That userinfo call is finally where we start getting into the actual Credit Nirvana code. It calls that API to obtain the preferred_username, which is then passed to their own login API. Since I had the users list, I modified this code to kill the userinfo API call and replace it with a hardcoded username. I picked one that had a high-level role:

This resulted in a successful login:

Notice how the Keycloak token is sent via the Authorization-Token header. This is the vulnerability – they did not check it! They chose to rely on the userName in the request body instead.

Inside the cnToken is a GUID that is used as a required CN-Token header in all subsequent requests. Unfortunately, that token + the fake ones did not result in the next API succeeding:

I found that if you removed the fake token headers, leaving only CN-Token, the request would go through!

And that was all I needed to do to gain full control. All APIs beyond this point were vulnerable to the same thing, so they all went through successfully.

By the way, those Keycloak patches up until the Credit Nirvana code can be used in other web apps. Another vulnerability to be disclosed in December could be exploited in a similar fashion.

The exposure nightmare

What kind of things were left exposed by this vulnerability? Buckle up, there’s a lot!

SlashRTC Admin

First off is admin-level access to SlashRTC – “a contact center and cloud communication provider offering AI-powered true Omnichannel platform.”

Accounts

Then we have the meat: all the accounts. Each one is linked to a person or business. Aditya Birla Capital Finance had more than 3.4 million:

Clicking into an account lets you view all the information about it:

Things get more sensitive when accounts have settlement data. In the cases, you can find all the sad stories and super-sensitive PII like ID cards and credit reports:

This credit report was 26 pages and contained all the person’s PII and lifetime account history:

There were more than 161k of these settlements:

Next up in the accounts are various types of documents you could download:

The Activity tab is where you can see everything that has happened with the account, like emails and text messages that were sent.

And finally, the legal notices. Any nasty/scary notices sent out are listed here.

If you are wondering if it was possible to cancel a debt, it’s unclear. There was no “Cancel Debt” button or similar that I saw.

Users & Orgs

Remember the more than 10k users exposed by the first API test? Those are managed here, and there is also a section to manage organizations. It is not just Aditya Birla Capital Finance staff, there are 433 debt collection agencies that are here to collect peoples’ debts. Talk about a team effort!

Digital Execution

This dashboard tracks all the digital communications, such as SMS, email, and WhatsApp messages sent to people about their loans. What is crazy about the below screenshots is that they are representative of only 1 month: May 2026 and yet there are millions and millions of records!

Receipts

We’re not done yet! There is plenty of sensitive information in receipts. There are 3.7 million receipts for May 2026 and 3.6 million for April 2026:

For each one you can click into it to view sensitive information, often including a picture of a check: Receipt Authorization lets you approve or reject them:

Activity Tracking

This is where we cross into the real world. In true dystopian fashion, apparently the debt collectors that go out to shake people down are tracked themselves. You can see where they went on a specific day, who they visited, and the outcome of the visit:

Call Logs

In May 2026, there were 214k calls made by their voice bot:

You could listen to all of them and view an AI summary:

Legal

In May 2026, 119k notices were sent out:

Among the user roles, there were some for lawyers. This is the dashboard they get to use to manage their legal cases. As admin, I could view all of them:

Other Clients/Instances

All of the above was captured from Aditya Birla Capital’s instance. If you look on the Credit Nirvana homepage, you will find a few other clients:

I located 2 other instances that were vulnerable to CVE-2026-110475:

Each instance is separate from each other, so that’s an additional 3.2 million accounts on top of the 3.4 million in Aditya Birla Capital for a total of 6.6 million.

Timeline #

Special thanks to India’s Computer Emergency Response Team (CERT-In) for working with me on this disclosure. Since this was a platform-level vulnerability affecting several clients, CVEs were issued.

Credit Nirvana / Perfios fixed the vulnerabilities the same day they became aware of them. You can’t ask for a better response than that. 🏆

  • May 30, 2026: Reported via CERT-In.
  • June 1, 2026: CERT-In wants a PoC video. One is created and sent.
  • June 2, 2026: CERT-In confirms they will take action. Later in the day, both vulnerabilities are fixed!
  • October 26, 2026: Vulnerability note, CVE-2026-110474 & CVE-2026-110475 issued. It is unclear why it took so long.
  • November 18, 2026: Published. Delayed due to various internet meltdowns. 💀
── more in #ai-agents 4 stories · sorted by recency
── more on @credit nirvana 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/the-ai-repo-man-expl…] indexed:0 read:10min 2026-11-18 · —