Researcher discloses AI debt-collection flaws that exposed 6.6 million accounts
The vulnerabilities also opened access to recorded calls and AI campaign controls. They were patched in June; malicious exploitation has not been established.
Loading page…
The vulnerabilities also opened access to recorded calls and AI campaign controls. They were patched in June; malicious exploitation has not been established.
Listen to this story
The exposed systems combined sensitive borrower histories—including credit records, legal documents and collection communications—with controls for managing AI agents and campaigns. Researcher Eaton Zveare attributed the reach to two separate authentication failures and estimated 6.6 million accounts and roughly 315,000 call recordings were accessible; the findings do not establish that attackers used the flaws. CreditNirvana fixed the bugs in June after notification, while public disclosure and the detailed account followed later.
The main platform’s deployments at Aditya Birla Capital Finance, BharatPe and International Asset Reconstruction Company exposed more than 3.4 million accounts at the first company and 3.2 million across the other two.
Call-center access included recordings, transcripts, AI-generated follow-up recommendations and the ability to manage agents or direct campaigns.
CERT-In rated the call-center flaw high severity and the main-platform flaw critical, warning that the latter could enable impersonation of administrators.
Two flaws in CreditNirvana’s AI debt-collection systems left millions of borrower accounts and hundreds of thousands of recorded calls accessible without a valid login, researcher Eaton Zveare disclosed on November 18. His findings describe more than a privacy failure: the weaknesses could let an intruder impersonate administrators or manage debt-collection AI agents.
Zveare puts the exposure at 6.6 million accounts and about 315,000 call recordings. CreditNirvana, owned by Indian fintech company Perfios, patched the bugs in June; they were publicly disclosed in October, ahead of this detailed account. The findings do not establish that malicious attackers exploited them.
The call-center flaw involved an API key, a credential that lets software access a service. The key was embedded in JavaScript sent to users’ browsers. Zveare says the service ignored the token returned by its Amazon Cognito login system. That let users bypass the visible sign-in step and use the exposed key to access the service’s functions.
The main collections platform failed differently. Its login service trusted a supplied username without checking the accompanying authentication token, according to Zveare. An unauthenticated endpoint also returned more than 10,000 users, giving him usernames to choose from. Together, those weaknesses let him establish an administrator session rather than merely look around a login page.
India’s Computer Emergency Response Team, CERT-In, rated the call-center vulnerability high severity and the main-platform flaw critical. As WIRED reported, CERT-In warned that the latter could allow impersonation of arbitrary users, including administrators, with access to read or change information under those users’ privileges.
The two systems exposed different data. The call-center recordings concerned Aditya Birla Capital Finance customers. Zveare identified vulnerable main-platform deployments at that company, BharatPe and International Asset Reconstruction Company. These were separate instances: more than 3.4 million accounts at Aditya Birla Capital Finance and another 3.2 million across the other two clients.
The main platform’s exposed material went well beyond names and contact details. It brought together financial histories, collection activity and documents used in disputes or legal proceedings:
In the call center, access extended from listening to calls and reading transcripts to reviewing AI-generated recommendations for follow-up actions. An intruder could also manage AI agents and direct campaigns. The exposure therefore covered both records of past interactions and tools for organizing future contact with borrowers.
Zveare credited CreditNirvana and Perfios with fixing both vulnerabilities the same day they became aware of them. His coordination with CERT-In included supplying a proof-of-concept video at the agency’s request. That response distinguishes the speed of the repair from the much longer interval before publication of the detailed findings.
CreditNirvana and Perfios did not respond to WIRED’s requests for comment. The technical account and exposure figures remain attributed to Zveare’s investigation; the companies supplied no response to that outlet explaining the failures or their impact on customers.
Loading discussion...
Join the conversation
Explain whether the added separation would be worth the inconvenience.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.