If a lender hands its borrowers' calls and credit files to an AI vendor, a hole in that vendor's login code becomes the lender's problem too. This one sat open for an unknown stretch, and nobody has said whether anyone else walked through it.
Researcher Eaton Zveare published details on Wednesday, November 18, of two vulnerabilities in CreditNirvana, an India-based, AI-driven debt collection platform. According to his write-up, the flaws exposed about 6.6 million borrower accounts and roughly 315,000 recorded calls, and an attacker needed no valid login to reach them.
The recordings came from customers of Aditya Birla Capital Finance, which alone accounted for more than 3.4 million of the exposed accounts. BharatPe and International Asset Reconstruction Company made up the other 3.2 million, according to the disclosure. Zveare says CreditNirvana, a subsidiary of Perfios, patched both bugs on June 2, three days after he reported them to India's CERT-In on May 30. The CVEs, CVE-2026-110474 and CVE-2026-110475, followed on October 26.
The two bugs were embarrassingly ordinary. In the agentic call center, Zveare says an API key was hardcoded in client-side JavaScript, and he could skip Amazon Cognito authentication by editing the page in his own browser. In the main platform, the server didn't validate Keycloak tokens at all. It trusted a username typed into the request body, so naming an administrator was enough to be treated as one.
Neither flaw needed an exploit kit. Anyone who opened the browser's developer tools could have found the first.
FBI and Google warn AI hacking tools like Artex now break into networks in hours Google's threat team named Artex, Cairn, HexStrike AI and PentAGI in a new report, while the FBI says Artex campaigns have hit at least ten sectors, including South Korean banks and ConnectWise customers. More than 1,600 control consoles sat exposed online in early November. - FBI warning about Artex AI hacking tools - AI agents breaking into networks within hours
What sat behind them is the unsettling part. The call center held transcripts, sentiment analysis and recommended dispositions for each of the 315,000 recordings, all from Aditya Birla Capital Finance customers. Superpower Daily, summarising the research, adds credit reports, emails and text messages to the list. Zveare also says the flaws could have let an intruder impersonate administrators or manage the collection AI agents themselves.
That last point deserves more attention than the headline count. Reading 315,000 calls is a privacy disaster. Controlling the agent that phones debtors is worse, because it puts an attacker's script in a collector's voice, talking to people already primed to pay whoever sounds official.
Who pays when the vendor slips #
The disclosure is narrow on a few points, and you should read it that way. It remains unclear how long the exposure lasted and whether anyone exploited it before the fix. Zveare's account describes a quick patch and says nothing of attackers, and we haven't seen a statement from the lenders. The vendor, Perfios and the three clients haven't, in what we found, said whether they notified affected borrowers.
Frankly, that silence is the story for anyone who buys these tools. The three clients didn't write the code, but their customers' calls and credit files were in it. A bank can outsource the collections workflow. It can't outsource the answer to a regulator or a borrower asking why a stranger could read their hardship call.
The pattern isn't new this year. Sears Home Services had nearly 3.7 million customer service records exposed through unsecured AI support bot databases, as SC Media and others reported earlier in 2026. An unsecured S3 bucket once held more than 55,000 recorded calls between loan workers and student borrowers, with Social Security numbers in the audio and paperwork, Cybernews found. Each time, the AI layer added volume, not a new kind of mistake. The mistake was a missing server-side check.
And volume is the point. A human call center keeps its audio scattered across agents' machines and a recording system nobody loves. An AI collections platform pulls the audio, the transcript, the credit file and the sentiment score into one searchable place, built to be queried quickly. One missing token check and the whole pile opens at once.
If you run or buy collections software, ask your vendor two things this week: where authentication is enforced, and how long it keeps recordings. Zveare's findings show that a platform can pass a sales demo with 315,000 calls stored and a login screen that was only ever decoration.
[Researchers find 1,465 exposed Artex AI hacking consoles open on the web](https://startupfortune.com/researchers-find-1465-exposed-artex-ai-hacking-consoles-open-on-the-web/)
Censys found 1,465 Artex consoles exposed online as of November 4, with 32 open directories leaking Windows lateral-movement tools, a sudo exploit and Linux implants that anyone could grab and reuse. - exposed Artex AI hacking consoles found on internet - Chinese AI penetration testing tool security breach discovery
Zveare reported the bugs on May 30. Public disclosure came 172 days later.
Also read: Meta staff flagged Muse AI reading Signal messages and mixing user data • GAO got ChatGPT, Claude, Gemini and Grok to plan US power grid attacks • Citi, PNC and Citizens tell customers they own any loss an AI agent causes
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.