All services

Services

AI & ML Testing

Assess whether model inputs can trigger unauthorized actions or expose protected data.

Discuss this service

How I test

I test the application around the model: identity, retrieval, connectors, and tool permissions. Findings require an observed security impact, such as unauthorized data access or an unintended tool action.

Untrusted input and retrieval

Direct and indirect prompt injection through user input, retrieved documents, websites, and other model-visible content.

Tools and agent actions

Tool permissions, approval boundaries, delegated credentials, and agent workflows. Test whether application controls enforce the user’s authority.

Data and tenant isolation

Retrieval permissions, memory, conversation state, connectors, and sensitive output across users and tenants.

What you receive

End-to-end test cases
Input, application state, tool behavior, and observed impact.
Boundary analysis
Enforcement gaps in application code, identity controls, and retrieval permissions.
Regression scenarios
Repeatable checks with observed variability and result limitations.