explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR: what Eaton demonstrated
  • Why agentic collections are a high-value target
  • Attack chain: from hardcoded key to full control
  • What data classes were exposed
  • What developers building agents should fix
  • How to protect collections and agent platforms
  • Collections agents versus coding agents: different blast radius
  • Keycloak and React SPAs: why client patches are not security
  • Disclosure timeline and what to watch next
  • Red-team checklist for agentic fintech UIs
  • Regulatory and industry context
  • Related reading
  • Why this belongs on the felony-bench thread
  • Related workshops and skills
← Back to blog

explainx / blog

Credit Nirvana CVE-2026-110474: How an Agentic Collections Stack Leaked Millions of Calls

Security, AI Agents, Vulnerabilities, Fintech, Debt Collection

Part of AI Security

Researcher Eaton disclosed CVE-2026-110474 and CVE-2026-110475 in Credit Nirvana on Nov 18, 2026, exposing agentic call logs and collections data for Aditya Birla Capital Finance.

Nov 18, 2026·8 min read·Yash Thakker
add explainx.ai
go deep
Credit Nirvana CVE-2026-110474: How an Agentic Collections Stack Leaked Millions of Calls

Security researcher Eaton published a full compromise walkthrough of Credit Nirvana's agentic debt-collection stack on November 18, 2026, assigning CVE-2026-110474 and CVE-2026-110475. The write-up shows how authentication could be skipped, Keycloak session handling could be bypassed in the browser, and backend APIs trusted the wrong headers, exposing agentic call tools and enterprise collections data tied to Aditya Birla Capital Finance.

This is not a theoretical agent prompt attack. It is production fintech software marketing AI agents for collections, with hundreds of thousands of call recordings reachable when identity checks fail.

Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

TL;DR: what Eaton demonstrated

table · 2 cols
SurfaceWhat Eaton reported
Agentic call centerCall Transcriber, disposition, post-call recommendations, call audit agents
Recordings~315,000 stored calls with playback
Core platformCollectionWorkz Credit Nirvana instance for Aditya Birla Capital Finance
Auth flawsHardcoded API key, unused Cognito token, 10k+ users via unauthenticated list API
IAMKeycloak client patches with fake JWTs; server trusted userName in body over token
CVEsCVE-2026-110474, CVE-2026-110475

Why agentic collections are a high-value target

Debt-collection software already holds names, account numbers, payment history, and call audio. Vendors now brand the same stack as agentic AI: real-time transcription, sentiment analysis, next-best-action recommendations, and proactive outreach campaigns before accounts go delinquent.

Eaton's screenshots and descriptions show those features as first-class modules, not future roadmap slides. An attacker who gains operator-grade access can replay calls, study negotiation tactics, and map which agents run automatically — intelligence useful for fraud, social engineering, and regulatory embarrassment.

The incident belongs in the same conversation as GAO tests showing models draft attack plans and agent breach trackers, but the exploit chain here is Web 101: missing server-side authorization on APIs behind a React single-page app.

Agent code review pair illustrating weak auth around Credit Nirvana agentic debt toolsAgent code review pair illustrating weak auth around Credit Nirvana agentic debt tools

Attack chain: from hardcoded key to full control

Enterprise pilot rollout metaphor for agentic collections UIs that skip server-side auth checksEnterprise pilot rollout metaphor for agentic collections UIs that skip server-side auth checks

Eaton's narrative starts at the agentic call center, where a hardcoded API key made the Amazon Cognito login unnecessary. The returned Cognito token was never validated server-side, so skipping login was enough to reach privileged features.

On the main CollectionWorkz deployment, Eaton traced React bundles, found a users-list API returning more than 10,000 accounts without authentication, and pivoted to Keycloak integration. By feeding syntactically valid but invalid JWTs and patching client-side refresh logic, the browser could be convinced a session might exist.

The critical backend mistake: Credit Nirvana's login API honored userName in the JSON body while sending a Keycloak token in Authorization-Token that was not verified. Eaton picked a high-privilege username from the leaked list, obtained a cnToken GUID, and discovered that subsequent requests worked when fake Keycloak headers were stripped and only CN-Token remained.

That pattern — trusting an application token while ignoring the enterprise IAM token — is how agent orchestrators and human portals get owned in the same request path.

What data classes were exposed

Eaton describes call recordings with transcripts, disposition analytics, campaign builders for AI-powered outreach, and an AI Agent Registry to browse agents inside the collection system. On the core platform, user tables, role metadata, and collections workflows across secured and unsecured products were in scope once APIs accepted the CN-Token.

Eaton notes Aditya Birla Capital Finance customers as impacted on the Agentic AI Call Center, and documents a far larger CollectionWorkz exposure on the primary platform: 6.6 million accounts across Aditya Birla Capital Finance, BharatPe, and International Asset Reconstruction Company (IARC), with ID cards, credit reports, legal documents, and call recordings in scope. Credit Nirvana and Perfios fixed both issues in June 2026 after CERT-In coordinated disclosure; the November 18 article is Eaton's public write-up, not a final statutory breach notice. Check issuer advisories and Indian regulatory filings for confirmed customer-notification timelines.

What developers building agents should fix

  1. Never embed long-lived API keys in front-end JavaScript. Any key shipped to browsers is public.
  2. Validate IAM tokens on the server for every route (Keycloak's token validation docs describe local JWT checks and introspection), including internal microservices behind your agent UI.
  3. Do not accept usernames from request bodies when a bearer token is present. Bind identity to verified claims only.
  4. Treat agent registries like admin consoles. They reveal tool names, prompts, and automation scope attackers need for follow-on abuse.
  5. Log and alert on unauthenticated list endpoints such as /users-list patterns Eaton highlights.

Teams shipping MCP-connected agents should compare this to MCP tool provider misbinding and Microsoft's PyPI credential stealer targeting agent frameworks — different entry points, same lesson: agents amplify impact when auth is weak.

How to protect collections and agent platforms

If you operate voice agents, collections SaaS, or internal agent hubs:

  • Run external pen tests that assume browser tampering, not only OWASP Top 10 scanners.
  • Segment call storage with per-tenant encryption and break-glass access logging.
  • Apply rate limits and anomaly detection on transcript download APIs.
  • Require step-up authentication before campaign launch or agent registry edits.

For runtime guardrails on autonomous tools — approving transfers, sending messages, or changing account state — teams can use platforms such as AgentBeam to block dangerous agent actions before they execute, alongside fixing the authorization bugs Eaton documented.

Collections agents versus coding agents: different blast radius

Many teams worry about Claude Code prompt injection or Codex sandbox escapes. Collections agents add telephony and regulated speech:

  • Outbound dialers can harass consumers if misconfigured.
  • Transcripts may include third-party bystanders on speakerphone.
  • Disposition models influence legal escalation paths.

Eaton's post is a reminder that the scariest agent failures may be missing auth on /users-list, not clever jailbreaks. When you prioritize controls, inventory admin routes before you tune system prompts.

Keycloak and React SPAs: why client patches are not security

Eaton documented disabling refresh handlers and short-circuiting userinfo calls inside the browser. That is powerful for proof of concept, and terrifying for defenders, because it shows how little the server cared.

Production fixes must assume attackers control the client:

table · 2 cols
Bad patternSafer pattern
Trust userName in JSON bodyBind identity to validated JWT claims server-side
Accept requests with only app tokenRequire IAM token + app token with matching subject
Ship API keys in bundlesUse short-lived tokens from backend exchange
Expose user enumeration APIsGate lists behind admin RBAC and rate limits

Credit Nirvana's CN-Token header behaving like a session cookie without binding to Keycloak matches failures seen in MCP tool misbinding where routers treat metadata as proof of identity.

Disclosure timeline and what to watch next

Eaton published on November 18, 2026 with CVE-2026-110474 and CVE-2026-110475 identifiers in the title and metadata. The author notes another Keycloak-related issue planned for December, suggesting coordinated disclosure cadence.

Customers should monitor:

  • Credit Nirvana or Aditya Birla Capital Finance security statements
  • CERT-In or sector regulators if notifications become public
  • CISA KEV listing if vendors assign CVSS and active exploitation proof

Until then, treat Eaton's article as technical primary source, not a final victim count.

Red-team checklist for agentic fintech UIs

Use this as an internal test plan:

  1. Download the production JavaScript bundle and grep for apiKey, cognito, and Bearer.
  2. Call every /users and /agents list route without cookies.
  3. Replay tokens from a low-privilege user against high-privilege routes.
  4. Remove headers one at a time to see which the API actually enforces.
  5. Attempt campaign creation without MFA if the product claims human approval.

These steps mirror how nginx2root HTTP/2 RCE and Java4Shell reached production — boring HTTP, enormous impact.

Regulatory and industry context

Agentic collections sit under fair debt, telemarketing, and financial privacy rules in multiple jurisdictions. A breach that exposes call audio and AI-generated dispositions can trigger contractual penalties even before class actions arrive.

The case also feeds felony-bench-style tracking of agents affecting real third parties: here the harm is data exposure, not an autonomous wire transfer, but collections agents can still initiate outbound contact when operators misuse restored access.

Related reading

  • MCP tool misbinding in agents
  • Microsoft Agent Framework PyPI stealer
  • GAO covert AI attack-plan tests
  • What is MCP?
  • AI agent security platforms

Why this belongs on the felony-bench thread

explainx.ai tracks agents that harm third parties on felony-bench. This incident is data exposure, not an autonomous funds transfer, but it shows agent branding on high-risk workflows without agent-grade authorization.

Security teams should log it beside GitHub and Microsoft's MDASH disclosure pipeline as a case where "agentic" marketing outpaced API maturity.

Related workshops and skills

If your team is shipping voice or text agents into regulated industries, pair technical fixes above with live agent workshops on permission models and eval harnesses before you enable outbound automation.

Eaton closes by noting that Keycloak client tricks may apply to other React apps facing a December disclosure. Even if you do not run Credit Nirvana, re-test your IAM integration any time you ship new agent dashboards to customer support or collections teams.

Technical details and CVE identifiers come from Eaton's November 18, 2026 article. Impact scope for specific customers may change as vendors publish advisories.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 18, 2026

AWS Bedrock AgentCore Identity: Private CA Support for Self-Hosted IdPs

On November 18, 2026 AWS documented private CA support in Amazon Bedrock AgentCore Identity, letting agents trust TLS from self-hosted identity providers via Secrets Manager or S3 trust anchors and VPC Lattice private endpoints.

Nov 18, 2026

Claude Gets Paranoid About Remote SSH: What Builders Saw on HN

Developers on Hacker News reported Claude refusing to finish a remote Ubuntu dual-boot job until the user created a local RTF file proving physical access to the control machine. The behavior looks like intentional friction for high-risk remote actions.

Nov 18, 2026

NGINX2Root: Five HTTP/2 Flaws, Two Critical RCE CVEs Patched Nov 18

On November 18, 2026 F5 shipped an NGINX security release for five CVEs collectively called nginx2root. The worst flaws are unauthenticated HTTP/2 remote code execution bugs that Xint, Wiz, Rapid7, and other teams found with AI-assisted review. Xint credits GPT 6.1 Sol and GPT-6 Luna for four of five issues in its public write-up.