Thursday, Nov 19 | --:--
Back to home

CreditNirvana AI Debt Collector's Two Flaws Exposed 6.6M Accounts, Researcher Says

Researcher Eaton Zveare says two flaws in CreditNirvana, an India-based AI debt-collection platform owned by Perfios, exposed about 6.6 million borrower accounts and about 315,000 call recordings from its agentic call centre, WIRED reported on November 18. India's CERT-In rated one flaw critical (CVSS 9.3) and the other high (8.7); Zveare says they were fixed on June 2. The counts are the researcher's, CreditNirvana and Perfios did not respond to WIRED, and no source we read reports that anyone else exploited the flaws.

5 min read Mumbai View as Markdown
A handmade editorial collage on torn ivory paper in cobalt and vermilion: a large reel of dark brown cassette tape with a cobalt hub on a plain grey plaster wall unspools a ribbon of tape through a jagged crack in the wall, the ribbon curling out into pale torn paper like daylight, with a few cobalt scraps and one tiny vermilion scrap; an official Credit Nirvana logo, with its hexagon mark, wordmark and 'A Perfios Company' tagline, is glued on a torn-paper clipping tucked under the reel's lower edge and the tape ribbon at bottom left, the full logo visible.

Neither flaw was an AI failure. One was a login service that did not properly verify who was asking; the other was a hardcoded key in client-side code. What made them serious is what an AI debt-collection platform keeps behind its login: borrowers' credit reports, ID cards, call recordings and the controls for the AI agents that phone them. Researcher Eaton Zveare says two flaws in CreditNirvana, an India-based platform owned by the fintech company Perfios, exposed about 6.6 million borrower accounts and about 315,000 call recordings, WIRED reported on November 18. India's CERT-In rated one flaw critical and the other high, and Zveare says both were fixed in June.

What the researcher says#

Zveare's write-up describes two separate problems. In CreditNirvana's Agentic AI Call Center, a hardcoded API key in client-side JavaScript gave access to about 315,000 call recordings from customers of Aditya Birla Capital Finance, along with the platform's post-call recommendations. He says it also let a visitor build and manage AI-powered campaigns and browse and manage AI agents.

In the main platform, called CollectionWorkz at Aditya Birla Capital, he says the login service did not verify the identity token it was sent and relied on a username in the request instead, which let him sign in as an administrator. He counts more than 3.4 million accounts at Aditya Birla Capital Finance and another 3.2 million at two other clients, BharatPe and International Asset Reconstruction Company, for the 6.6 million total. He says the accounts held ID cards, credit reports and legal documents, with millions of emails and texts, and that the platform could track where human collection agents went. WIRED quotes CERT-In's description of the critical flaw: "Successful exploitation of this vulnerability could allow the attacker to enumerate platform users, establish sessions as arbitrary users including administrative users, access and modify sensitive information with the privileges of the impersonated user."

Fixes and timeline#

CERT-In issued a combined bulletin, CIVN-2026-0523, on October 26 with two CVEs: CVE-2026-110475, rated critical at 9.3, for the CollectionWorkz login flaw, and CVE-2026-110474, rated high at 8.7, for the call-centre key. The fixes are CollectionWorkz 4.8.2 and Agentic AI Call Centre 2.0.0, according to Gloqo's summary of the advisory.

Zveare's timeline has him reporting through CERT-In on May 30, a fix on June 2, the CERT-In note on October 26 and his post on November 18. He says the vendor fixed both the same day it became aware, and that the flaws may have been present for years, which is his inference. CreditNirvana and Perfios did not return WIRED's requests for comment.

Claims and checks#

The scale comes from one person. The 6.6 million accounts, the 315,000 recordings and the client names are Zveare's, and WIRED does not report an independent count. CERT-In's severity ratings are a separate check on how serious the flaws were, as reported by Gloqo and by an index of the CVE record, though not on how much data was reachable.

The word "exposed" needs care. WIRED's body says the vulnerabilities "would have allowed" attackers to reach the data. Gloqo says the CERT-In advisory and CVE records it reviewed do not report in-the-wild exploitation and do not establish that any customer data was accessed, and the CVE index lists no known exploitation. That is consistent with exposure, not evidence of theft. Another AI-agent security failure we covered on November 18 was the poisoned Microsoft Agent Framework releases, which, like this case, turned on ordinary software security rather than model behaviour.

Limits#

  • The account counts, client names and data types are the researcher's. CreditNirvana and Perfios did not respond to WIRED, and we found no vendor statement.
  • No source we read reports exploitation by anyone other than the researcher, but that does not establish that none occurred.
  • Dates differ in detail: WIRED says the bugs were patched in June and disclosed in October, while Zveare's timeline gives June 2 for the fix and November 18 for his full write-up.
  • We did not open CERT-In's own records; severity and version details come through Gloqo's summary and a CVE index.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading