Security testing of LLM applications, chatbots and AI agents
More and more companies integrate large language models (LLMs) into their products and internal processes: customer-facing chatbots, assistants that search internal documents, and AI agents that independently call tools, APIs and databases. These systems introduce a new class of vulnerabilities that classic penetration tests do not cover. Our AI pentest examines how your AI application can be manipulated, what data it can leak and what actions an attacker can trigger through it. We follow established frameworks such as the OWASP Top 10 for LLM Applications and MITRE ATLAS.
Motivation
An LLM does not reliably distinguish between instructions and data. Every piece of text the model processes, whether it is a user message, an uploaded document, an email or a web page, can contain instructions that change its behavior. If the model has access to sensitive data or can perform actions, a successful prompt injection can lead to data leaks, unauthorized transactions or the compromise of connected systems. Because AI features are often built and shipped quickly, these risks are frequently overlooked. A targeted security test shows you where your AI system can be abused before attackers find out.
What we offer
At rasotec, we test your AI systems from the perspective of a real attacker. Depending on your application, our tests cover the following areas:
- Prompt injection and jailbreaks: We attempt to override the system instructions of your model, both directly through user input and indirectly through documents, emails, websites or other content the model processes.
- Sensitive data disclosure: We check whether system prompts, internal documents, personal data or data of other users and tenants can be extracted from the model or its knowledge base.
- RAG and knowledge bases: We analyze retrieval-augmented generation (RAG) pipelines for missing access controls, data poisoning and the leakage of documents the user should not be able to see.
- AI agents and tool use: We examine which tools, APIs and permissions your agents have and whether an attacker can make them perform unintended actions. This includes integrations via the Model Context Protocol (MCP).
- Insecure output handling: We test whether model output is passed unchecked to browsers, shells, databases or other systems, which can result in classic vulnerabilities such as cross-site scripting (XSS), SQL injection or server-side request forgery (SSRF).
- Surrounding infrastructure: We review the APIs, authentication, rate limits and cost controls around your model to prevent abuse, unauthorized access and unexpectedly high usage costs.
Our security experts combine manual testing with automated attack techniques to obtain a comprehensive picture of the security of your AI application.
After the test, you will receive a detailed report that documents the vulnerabilities found with reproducible examples and recommends specific measures such as input and output filtering, privilege reduction and architectural safeguards.
Protect your company and your customers with a professional AI security test from rasotec and use the potential of AI without opening new doors for attackers.