Representative engagement This case study describes a typical engagement of this kind rather than a specific client. Figures are representative of the work, not a guarantee of results.
AI security review · Financial services
Eleven unapproved AI tools, one customer-data leak, two weeks
A 60-person payments company needed to answer a bank partner's AI questionnaire and had no clear picture of how AI was being used across the business. The review found the picture, fixed what mattered, and produced the evidence the bank asked for.
Results
At the end of the two-week review and the four-week remediation that followed.
The problem
The company's bank partner had added an AI section to its annual due-diligence questionnaire: which AI tools were in use, what data they received, and how the company's own customer-support assistant was protected against manipulation. Nobody could answer with confidence.
Staff were using a mix of personal ChatGPT accounts, browser extensions and AI features inside existing SaaS tools. The support assistant, built the previous year on a retrieval system over the help desk, had never been tested for prompt injection.
The managed service provider looked after devices and identities but had no view of the AI layer.
What we did
A two-week review, then remediation with the client's own engineers and their MSP.
Week one: discovery
Identity logs, network data and expense records revealed eleven AI tools in use, three with company data under consumer terms that permitted training. A data-flow map showed every path to a model provider.
Week two: testing
Indirect prompt injection through a customer message caused the support assistant to return another customer's transaction details. Agent permissions and logging of personal information were also flagged.
Policy and mapping
An AI use policy with three approved tools, mapped to the Essential Eight and the Australian Privacy Principles, including the automated-decision disclosure due in December 2026.
Remediation
The assistant was fixed by the client's engineers using our plan: tenant-scoped retrieval, output validation and a guardrail model. The MSP enforced the approved tools. All staff completed the 90-minute Safe AI at work session.
Stack
What made the difference
The shadow-AI count surprised the leadership team more than the assistant flaw did. Most of the eleven tools were adopted with good intentions by people who had no approved alternative.
Testing against the real assistant turned an abstract risk into a reproducible demonstration, which is what moved remediation to the top of the engineering backlog.
Working with the MSP rather than around it meant the fixes were enforced in the tools the company already paid for.