LLM Application Security Testing

Treat prompts and context as untrusted input

A user, uploaded document, webpage or retrieved record may contain instructions that compete with the application's intended rules. We test whether those instructions can influence behaviour or override safeguards.

Prevent sensitive information disclosure

Models can reveal information from prompts, conversation history, retrieved content or connected systems when access controls and data boundaries are weak.

Test retrieval and embedding security

RAG systems introduce vector stores, document ingestion and retrieval logic. We assess whether content can be poisoned, retrieved across the wrong boundary or used to manipulate the model.

Handle model outputs safely

LLM output may be rendered in a browser, passed to an API or used as an instruction by another system. We test whether unsafe output can lead to injection, code execution or unauthorised actions.

Understand where guardrails break down

Controls that work for a simple prompt may fail across different languages, roles, conversation lengths or chained instructions. Manual testing explores those edge cases.

Control excessive or abusive use

Without effective limits, attackers may consume disproportionate resources, degrade the service or increase model costs. We assess the controls intended to prevent that.