On November 18, 2026, Eaton Zveare published The AI Repo Man—a breakdown of two Credit Nirvana flaws that turned an agentic debt-collection stack into full administrative takeover. CERT-In CIVN-2026-0523 assigns CVE-2026-110474 (client-side API key misuse on the Agentic AI Call Center) and CVE-2026-110475 (critical authentication failure on CollectionWorkz). Vendor fixed builds start at 4.8.2 per CERT-In.
If you build or deploy agent registries, voice bots, or campaign orchestrators—even locally—the same mistakes appear when coding agents paste “temporary” API keys into React and wire Keycloak only on the login page.
Image: Eaton Zveare
TL;DR
| Question | Answer |
|---|---|
| Am I affected? | Only if you operate CollectionWorkz before 4.8.2 or the vulnerable call-center SPA |
| Exploitation class | Hardcoded key + fake Keycloak session → admin APIs |
| Data at risk | Accounts, ID scans, credit reports, call audio, AI summaries, legal notices |
| Agent ops risk | Campaign builders and AI agent registry were reachable |
| Fixed version | 4.8.2+ per CERT-In |
| Coding-agent lesson | Do not let agents commit static API keys or skip server auth |
What Credit Nirvana shipped
Credit Nirvana (Perfios) markets AI-assisted collections to Indian banks and NBFCs. Eaton’s research targets two surfaces:
- Agentic AI Call Center — Amazon Cognito login UI, but backend calls authenticated via a hardcoded API key in JavaScript. Features include Call Transcriber, Disposition, Post Call Action Recommender, Call Audit, Campaigns, and an AI Agent Registry.
- CollectionWorkz primary platform — Keycloak IAM for staff, React front end, SlashRTC integration, millions of accounts with settlements, credit reports, receipts, and field-agent GPS tracking.
Clients named in the disclosure include Aditya Birla Capital Finance, BharatPe, and International Asset Reconstruction Company (IARC).
Attack chain A: Agentic AI Call Center (CVE-2026-110474)
From Eaton’s write-up and CVE-2026-110474 summary:
- Researcher inspects the SPA; Cognito sign-in is cosmetic for API authorization.
- A hardcoded API key in client code enables backend access regardless of Cognito tokens.
- By commenting out the sign-in API call and substituting admin-shaped fields, any email/password satisfies the UI.
- Attacker reaches 315k stored call recordings, AI disposition outputs, and tooling to configure outbound AI campaigns.
- AI Agent Registry lists agents available to the collection system—administration plane for voice bots, not chat toys.
Impact highlight: Aditya Birla Capital Finance customers in the call-center deployment; ability to listen, read transcripts, and manage AI-powered campaigns.
This is a textbook hard-coded API key pattern (CWE-321 per CERT-In) amplified by agentic features: the same key that powers “monitor campaigns in real time” becomes a remote admin lever.
Attack chain B: CollectionWorkz / Keycloak (CVE-2026-110475)
CVE-2026-110475 describes improper authentication at the platform login service: attackers manipulate requests without a valid session.
Eaton’s step-by-step (primary source):
users-listAPI returns 10k+ users with no authentication—recon at scale.- Front-end Keycloak flow redirects unauthenticated users to login.
- Researcher patches JavaScript to supply syntactically valid JWT-shaped strings in storage, disabling refresh and userinfo calls that would fail.
- Instead of calling userinfo, code hardcodes a high-privilege
preferred_usernamefrom the leaked user list. - Credit Nirvana’s login API trusts
userNamein the body, sending a Keycloak token inAuthorization-Tokenthat the server does not validate. - Successful login yields
cnTokenused asCN-Tokenheader; attacker strips fake Keycloak headers and all subsequent APIs succeed.
Exposure scale (researcher): 3.4M+ accounts on one instance, 161k settlements with ID cards and 26-page TransUnion CIBIL credit reports, 3.7M receipts in a single month, 214k voice-bot calls in May 2026 with AI summaries, legal case dashboards, and SlashRTC admin.
Two additional client instances (BharatPe, IARC) brought the total to 6.6 million accounts.
Timeline and official CVE metadata
| Date | Event |
|---|---|
| May 30, 2026 | Reported via CERT-In |
| June 1–2, 2026 | PoC video; vendor fixes both issues |
| Oct 26, 2026 | CIVN-2026-0523, CVEs published |
| Nov 18, 2026 | Eaton public write-up |
CERT-In lists upgrade to 4.8.2 for CollectionWorkz. Eaton notes same-day vendor response after coordination.
What this means if you run AI coding agents
This incident is not about prompt injection in a coding CLI. It is about product security for agentic operations—the same class of system your team might prototype with LangGraph, n8n, or a custom MCP voice tool:
| Mistake | Credit Nirvana manifestation | Agent-team guardrail |
|---|---|---|
| Secrets in front-end bundles | Hardcoded API key | Reject PRs where agents add VITE_* or NEXT_PUBLIC_* service keys; use Beam PreToolUse hooks to flag risky shell commands before they run |
| IdP theater | Keycloak UI with unvalidated tokens | Require server middleware tests that 401 without valid OIDC claims—agents should not disable those tests to “fix CI” |
| Admin APIs without session bind | userName body trust | Integration tests that swap body username while keeping invalid token—must fail |
| Agent registry as live attack surface | Registry + campaign APIs | Treat registry write like production deploy—human review, audit log |
Beam scans skills/MCP configs and flags risky shell on developer machines; it does not audit remote SaaS like Credit Nirvana. The actionable mirror is your CollectionWorkz-like internal tools agents scaffold.
How Beam handles overlapping observability (for your own stacks)
If you operate self-hosted agent orchestration with similar campaign or registry APIs, Beam’s local-first model fits developer-side control: on-device audit trails for what Claude Code and Codex changed in repos, heuristic MCP/skill scans before install, and hook-based deny for destructive commands. That does not replace server-side auth Eaton proved missing—it complements it by catching agents that embed keys into your front ends during rapid prototyping. Limits: heuristic, not semantic malware analysis; fail-open if policy cache is missing.
Building safer agentic call flows (checklist)
- No long-lived keys in browser code—use short-lived tokens from a backend-for-frontend.
- Validate OIDC tokens on every API, including internal microservices agents generate.
- Separate “read call metadata” from “launch campaign” with distinct scopes—registry APIs are tier-0.
- Log agent actions (model ID, tool, operator) on par with human admins—for patterns see KPMG parallel agent audit.
- Red-team with browser devtools tampering exactly as Eaton did—agents will ship interceptors in React during refactors.
Disclosure lessons for agent-security readers
Eaton explicitly notes Keycloak front-end patches may generalize to other web apps with a December disclosure pending. Teams maintaining agent admin consoles should schedule auth reviews when that lands—not only npm patches.
The AI Repo Man framing is deliberate: agentic collections automate coercive workflows. Compromise is not just data theft—it is weaponized outbound agents with pre-built campaign templates.
Agentic modules named in the disclosure
Eaton’s walkthrough is useful because it lists production agent roles, not marketing labels:
| Module | Stated purpose (vendor UI) | Attacker capability on vulnerable build |
|---|---|---|
| Call Transcriber Agent | Real-time transcription, language detection, sentiment | Browse and listen to stored calls |
| Disposition Agent | Post-call disposition insights | Read model outputs tied to borrowers |
| Post Call Action Recommender | Next-best-action after calls | See recommendations that drive collector actions |
| Call Audit Agent | Audit agentic AI calls | Review automated call decisions |
| Campaigns | Proactive outreach before delinquency | Configure AI-powered dialing/messaging campaigns |
| AI Agent Registry | Browse/manage collection agents | Inventory and potentially alter agent definitions |
That registry is the closest parallel to how internal platform teams expose tool lists to Claude Code or Copilot—except here compromise leaks voice data and legal workflows, not source code.
Official CVE boundaries (CERT-In)
CVE-2026-110474 is scoped to the Agentic AI Call Centre product, CWE-321 hard-coded key, CVSS 4.0 8.7, fixed at version 2.0.0 per CERT-In solution text.
CVE-2026-110475 is scoped to CollectionWorkz, CWE-1390 weak authentication, CVSS 4.0 9.3, fixed at 4.8.2.
The same CERT-In note CIVN-2026-0523 ties both CVEs to one coordinated report. Eaton’s narrative shows defense failures at two layers—SPA secret leakage and API trust of usernames—that must be patched independently even though one researcher chain exploited both.
Why Cognito plus a static key fails agent products
Amazon Cognito can be a sound control when tokens gate every API. Credit Nirvana’s call-center app treated Cognito as UX chrome: commenting out the sign-in call still left backend authorization to a static key embedded beside minified React bundles—trivially extractable with browser devtools or curl on .js assets.
Teams using agents to “add Cognito in an afternoon” should enforce:
- Server-side
Authorizationon every route, including GraphQL and WebSocket upgrades used by realtime agent dashboards. - CI secret scanning on front-end bundles (
grepforsk-,AKIA, long hex keys) on every agent PR. - Separate roles for campaign authors versus call listeners so one key cannot do both.
Field operations and non-AI data still mattered
Even ignoring the AI surfaces, CollectionWorkz exposed GPS tracks of field collectors, receipt photos including checks, and legal case dashboards for counsel roles. Agentic AI did not create that sensitivity—weak API auth amplified it. Security programs evaluating “agent risk” should still prioritize classic IDOR and authZ on the non-model paths agents integrate with first.
Vendor response and verification for third parties
Eaton credits same-day fixes after CERT-In coordination (June 2, 2026). Public customers should not rely on that timeline retrospectively—verify CollectionWorkz 4.8.2+ and Call Centre 2.0.0+ on any managed instance, request pen-test evidence if a vendor hosts Perfios/Credit Nirvana for you, and rotate integration secrets issued before June 2026. Indian regulated lenders integrating Perfios APIs should map this disclosure to their vendor risk register even when they never exposed CollectionWorkz URLs directly—subsidiary SaaS auth failures propagate through B2B data feeds the same way a compromised MCP server poisons every downstream agent session.
Related reading
- Microsoft agent-framework PyPI typosquat
- Token rehearsal and browser agent identity
- CyberStrikeAI approval gates
- Darktrace Claude memory store C2 pattern
- What is AI agent monitoring?
- November 2026 attacks timeline
Frequently asked questions
What does creditnirvana do?
It sells collections workflow software—digital communications, legal notices, field tracking, and newer agentic AI call-center automation—for Indian lenders. CollectionWorkz is the enterprise hub; the call-center product focuses on AI transcription, disposition, and campaigns.
What are CVE-2026-110474 and CVE-2026-110475?
110474 is the hardcoded / mishandled API key on the Agentic AI Call Center (High). 110475 is weak authentication allowing arbitrary admin sessions on CollectionWorkz (Critical, CVSS 4.0 9.3 per CERT-In).
Were customers still vulnerable in November 2026?
Public write-up is disclosure education; vendor fixed in June 2026. If you run private instances, confirm version ≥ 4.8.2 and rotate API keys issued before the fix.
Why link this to coding agents?
Teams use agents to scaffold React admin UIs and auth interceptors quickly—exactly the code Eaton modified. The failure mode is skipping server validation, not model misalignment.
Could attackers launch AI collection calls?
On vulnerable builds, researchers accessed campaign builders and registry paths—sufficient to reconfigure outbound automation for the affected tenant.
What should I search in my own repos?
grep -R "apiKey\|API_KEY" src/ in SPAs, and audit axios/fetch interceptors agents added for “local dev bypass.”
Sources
- Eaton Zveare — The AI Repo Man (November 18, 2026)
- CERT-In CIVN-2026-0523
- CVE-2026-110475 (CERT-In CNA record)
CVE metadata and timeline reflect Eaton’s publication and CERT-In records as of November 18, 2026 UTC.

