Applications

AI/LLM Applications

Large Language Models (LLMs) enable building applications of the future, from intelligent document processing and chatbots to custom AI agents. Test your LLM-based application for vulnerabilities.

Scope of the pentest

In this penetration test, our ethical hackers examine your LLM application for security vulnerabilities and configuration errors.
The test can be conducted on premises or remotely.

Exemplary test objects:

LLM-Chatbot

Typical AI-based chatbot systems with knowledge base via RAG pipeline

Document Analysis

AI applications with OCR and document processing

AI Content Creation

AI systems for dynamic content creation (e.g., images, text, etc.)

Custom App

You have developed your own custom AI application?

8.5%

8,5% of all prompts contain sensitive data ¹

45.77%

45,77% of these prompts contain customer data ¹

Penetration test of AI/LLM-Applications

Our approach

The AI penetration test presented here includes a comprehensive security analysis of your AI applications, particularly systems based on large language models (LLMs). The focus is on identifying risks arising from the interaction between the model, user input, connected data sources, and integrated systems. The scope of the test can be specifically limited to an application, API, or infrastructure component of your choosing.

We test the AI system both at the infrastructure level (all accessible interfaces of the target application) and at the application level using AI red teaming. In doing so, we specifically simulate malicious prompts and manipulated inputs to uncover potential vulnerabilities in the model’s behavior as well as in the processing and transmission of information.

All tests include a structured analysis of potential vulnerabilities along the entire processing chain – from input interfaces and APIs to connected systems and services. In addition, an in-depth analysis of the model’s behavior is conducted at the application level under realistic attack conditions. This includes prompt injection, data exfiltration, model manipulation, output filtering bypass, and the verification of access controls and system boundaries. The testing is conducted in accordance with established security guidelines such as the OWASP Top 10 for LLM applications and the current penetration testing guidelines for large-language models from the German Expertenkreis KI-Sicherheit.