Application and Software Security
AI and LLM Security
Organizations connect AI assistants to customer records and business workflows. The risk mainly arises in these connections, not in the model itself. We test the assistant together with the data and tools it accesses. We show under which conditions the model can be steered beyond its authority.
AI and LLM Security
Large language models cannot always distinguish instructions from data. A document they read may also contain instructions. Data protection law requires high-risk systems to be resilient against attempts to alter their outputs by exploiting vulnerabilities.
The need
When organizations need it
- Before a customer service assistant is given access to order and customer records.
- When documents with different access levels are combined in a single knowledge base.
- Before an AI agent is allowed to initiate a payment step.
- When an enterprise customer requests an independent test report on an AI feature in your product.
Questions
What this work answers
- Can an instruction hidden in an email or document steer our assistant?
- Can a user access another customer's data through the model?
- Can our agents perform an irreversible action without approval?
- Does our application execute model output without validation?
AI and LLM Security
What we do
-
LLM Application Security Testing
We test the assistant together with its interface, permission structure and backend services.
-
Prompt Injection Assessment
We test instructions typed by users and instructions hidden in documents the model reads. We measure whether the system's behavior changes.
-
AI Agent Permission and Tool Security
We test whether tool permissions and approval steps prevent the action when the agent is misdirected.
-
RAG and Knowledge Source Security
We check whether document permissions are preserved in the model's responses. We also assess the trustworthiness of externally added content.
-
Sensitive Data and Tenant Isolation
We test whether one customer's data can be accessed by another through conversation history or caches.
-
Model Output Handling Security
We check whether model output is executed without validation in a browser or database query.
-
AI Supply Chain Security
We assess the origin of the models, plugins and libraries you use, and whether they have been modified.
-
AI Threat Modeling
Before the system goes live, we map abuse scenarios. We design permission and approval controls against them with you.
Scope
The test covers the system as seen by users, for every user role. The model, knowledge sources and tools are all in scope.
How we work
First, we identify every source the model reads and every tool it can call. Instructions can enter the system through these routes. We test scenarios along these routes using test accounts you provide. The model does not always give the same answer to the same input. Therefore, each scenario is repeated several times.
Deliverables
What you receive
For each scenario: how the model was steered, success rate, which data or action was reached, and controls to apply outside the model in the permission and approval layers.
NoteRestrictions added to the system prompt alone do not constitute a security control. Our recommendations focus on layers that stop the action even if the model is misdirected.
Related services
Schedule AI Security Meeting
Tell us your objective and critical systems. We will settle the right scope in the first meeting.
Schedule a Meeting