Case study - A full audit of a multi-tenant voice and chat AI platform

Overview

The platform answers phone calls and website chats for small businesses with AI agents that can book appointments, call people back and write to CRMs. Every business is a tenant in one shared database.

The audit covered about 140 API routes, 54 database tables, 21 functions, six storage buckets, the voice worker, telephony, payments and the AI providers.

What we found

Tenant isolation held: a live two tenant test against all 54 tables let no read, insert, update or delete cross between customers.

Two high findings were exploitable from outside. A script could bill customers for chat resolutions that never happened, and the voice agent could be talked into calling any number a caller dictated. Both were fixed and verified the same night.

Third party credentials were stored in plain text and readable from the browser session. All 49 were sealed with AES-256-GCM and moved out of the browser's reach.

What was delivered

A full report with 105 findings mapped to SOC 2, ISO 27001 and the OWASP LLM Top 10, a SOC 2 control matrix, an ISO 27001 Statement of Applicability, ten policies, risk and vendor registers, and a restore drill with zero mismatches across 11,562 rows.