AI & LLM SECURITY TESTING: Test the boundaries around your AI application.
AI and LLM security testing examines how an AI-enabled application handles untrusted instructions, sensitive information and connected tools. The assessment challenges the boundaries around the model, including retrieval and action-taking workflows, to establish how someone could influence its behaviour or gain access beyond the intended use.
- Discipline
- Cyber Defense
- Engagement type
- Assessment
More access gives a model's output greater consequences.
A chatbot answering general questions has a different risk profile from an agent that retrieves customer records or changes business systems. Testing should follow the application's actual access and authority, including content it retrieves from documents, websites or other external sources.
Common triggers
adding retrieval over private data; connecting tools or agents to live workflows; introducing external documents or user-generated content; releasing an AI feature to customers.
Trace inputs through data access and actions.
Untrusted instructions
Direct prompts and instructions embedded in retrieved or external content.
Information boundaries
Whether users or sessions can obtain material beyond their permitted access.
Tool permissions
The actions available to the application and the controls around consequential changes.
Output handling
How generated content is used by browsers, APIs and other downstream systems.
Abuse controls
Relevant restrictions on repeated requests, resource use and costly or sensitive operations.
Start with the use case and its failure conditions.
We map the agreed components, data sources and tools, then define scenarios that would create an unacceptable result. Findings record the conditions, observed behaviour and consequences, with guidance on controls in the application and surrounding services. The evaluated model, configuration and test conditions remain identifiable so that later changes can be considered in context.
What you receive
The scenarios, components and configuration examined.
Relevant input, retrieval or tool-interaction evidence showing observed behaviour.
Recommended application controls and follow-up checks tied to the failure conditions.
Is this only prompt testing?
No. Prompt behaviour matters, but system security also depends on retrieval permissions, tool access, output handling and the surrounding application.
Can testing prove prompt injection is impossible?
No. It provides evidence from the scenarios examined. Effective controls should also restrict what the application can access and do if model behaviour is manipulated.
Does this include bias, legal compliance or AI governance?
These involve different evaluation criteria and responsibilities. This service focuses on agreed security risks; broader governance requirements should be scoped separately.
What can your AI application read or change?
Tell us its purpose, model integrations, connected data and available tools. Those relationships determine the assessment.
Related services: API Security Assessment and ISO 42001.