AI Security

How a threat model mapped the AI risks in a multi-cloud platform at design stage

Client
Large financial services company
Service
AI/ML threat modelling
Scope
30+ components across AWS and Azure
Remediation
All recommendations implemented
The challenge

What did the client need?

A large financial services company had built an AI and machine learning platform for document processing, customer analytics and generative AI question answering. It spanned more than 30 components across AWS and Azure, and traditional assessments did not cover its AI-specific risks.

Engagement details

Client
Large financial services company
Platform
Generative AI service with RAG pipelines, plus an ML model hub
Scope
30+ components across AWS and Azure, and the data flows between them
Method
NIST SP 800-30, the NVIDIA AI Red Team framework and STRIDE
Input
Interviews with 8 stakeholders and the platform’s documentation
Remediation
All recommendations implemented by the client
Delivered by
ioSENTRIX AI security team
Our approach

How did ioSENTRIX approach it?

Our team interviewed the people who built and ran the platform, mapped every component, trust zone and data flow across both clouds, then worked through how an attacker would use each one.

What we found

What did ioSENTRIX find?

Prompt injection, missing access control in retrieval and open inference paths, found on paper first.

HighThreat 1

Prompts could steer the AI into misusing its access

Both direct and indirect prompt injection were possible: instructions placed in a user’s input, or hidden in a document the system retrieves, could change what the language model does.

Business impact. The platform could be pushed to reveal or act on sensitive financial data and customer information.

HighThreat 2

Retrieval did not check who was asking

Role-based access control was not enforced in the retrieval-augmented generation (RAG) pipelines, and some ML inference paths were unprotected. The model also found data exposure risks in logging, storage and document uploads.

Business impact. A user could receive answers drawn from documents their role should never see.

Two threat scenarios the model ranked High
Scenario 1 · prompt injection · High
  1. 1Crafted inputdirect or in a file
  2. 2Model follows itinstructions injected
  3. 3Misused accessdata or actions
Scenario 2 · retrieval access · High
  1. 1Any user askswhatever their role
  2. 2Retrieval runsno RBAC enforced
  3. 3Answer returnedrestricted content
Outcome: Data exposure (financial data and customer PII)
The impact

What changed as a result?

ioSENTRIX built a threat model of the whole platform in five phases. It identified more than ten threats, five of them High, including prompt injection, missing role-based access control in retrieval pipelines and unprotected inference paths, each with a fix. The client implemented every recommendation.

30+
components mapped across two clouds
5
High-severity threats identified at design stage
10+
threats, each with a specific fix

How many threats, and how severe?

10+ threats in total, including 5 High.

  • High 5
  • Moderate
  • Low
  • Very Low

What changed after the assessment?

What the client implemented
AreaThreat foundWhat the client implemented
Prompt injectionDirect and indirect injection could steer the modelDefences against direct and indirect injection
Retrieval accessNo role-based access control in RAG pipelinesAccess control for retrieval and inference paths
Data handlingExposure in logging, storage and document uploadsFixes for logging, storage and uploads

✓ Verification: All recommendations implemented by the client.

Key insight

The models were new. The questions were old: who can ask, what can they reach, and who checks?

Lessons

Lessons for teams building on AI

01

AI needs its own threat model

Prompt injection ranked High, yet earlier assessments did not look for it. Treat AI as a new attack surface.

Threat 1 · OWASP Top 10 for LLM Applications (2025), LLM01: Prompt Injection
02

Enforce access at retrieval

Without RBAC in the RAG pipeline, the model returned what the pipeline handed it, whoever was asking.

Threat 2 · OWASP Top 10 for LLM Applications (2025), LLM08: Vector and Embedding Weaknesses
03

Map across cloud boundaries

Across 30+ components in AWS and Azure, trust broke where data crossed providers.

Approach
  1. Severity rated on a five-level likelihood and impact matrix based on NIST SP 800-30 Rev. 1. Levels shown: High, Moderate, Low and Very Low.
Next step

Building on generative AI? See a sample AI threat model or book a strategy session before your platform scales.

Service
AI security assessment
Explore the service

Details in this case study are anonymised and generalised. Methods that could harm other products are withheld.

FAQ

Frequently asked questions

What did ioSENTRIX do in this case study?

A large financial services company had built an AI and machine learning platform for document processing, customer analytics and generative AI question answering. It spanned more than 30 components across AWS and Azure, and traditional assessments did not cover its AI-specific risks.

What was the most serious issue?

Prompts could steer the AI into misusing its access. The platform could be pushed to reveal or act on sensitive financial data and customer information.

How many threats were there, and how severe?

10+ threats in total, including 5 High.

What happened after the engagement?

The client implemented every recommendation, using a reusable framework to assess each new AI component.

Which ioSENTRIX service does this case study cover?

AI security assessment, delivered by the ioSENTRIX AI security team.

Which standards does this align with?

OWASP Top 10 for LLM Applications (2025), LLM01: Prompt Injection; OWASP Top 10 for LLM Applications (2025), LLM08: Vector and Embedding Weaknesses. These are cited as alignment guidance, not a compliance claim.

How to get started

Ready to strengthen your security? Fill out our quick form, and a cybersecurity expert will reach out to discuss your needs and next steps.
DecorativeDecorative