AI Penetration Tests - Trovent Security GmbH

For whom are AI pentests relevant?

When artificial intelligence becomes part of your infrastructure, it also becomes part of your attack surface! Targeted AI and LLM penetration tests can uncover unknown vulnerabilities in your systems. Such specialized AI pentesting is particularly relevant for companies that…

… operate an AI-powered chatbot with access to internal documents or customer data.

… require documented proof of security testing under NIS2, DORA, or the EU AI Act.

… deploy or plan AI agents with tool access (email, ticketing systems, databases, payment processes).

… operate in the critical infrastructure sector or in regulated industries (e.g., banks, insurance companies, hospitals).

Get in touch with us!

Together with you, we assess whether and to what extent an AI security test makes sense for your systems.

Why traditional pentests are not sufficient here

AI-powered applications introduce attack surfaces that are not covered by traditional web, mobile, or infrastructure penetration tests. These include, among others:

Prompt injection

Attackers inject instructions via user input or manipulated documents that the model executes, contrary to its intended function.

Indirect prompt injections

Malicious content does not enter your IT environment via the chat, but rather through documents, tickets, or websites that are automatically retrieved by the system.

Excessive agency

Over time, AI agents gain access to more tools (sending emails, modifying databases, initiating payments) than were originally intended and tested.

Data leaks via RAG pipelines

Access rights to the underlying knowledge base are not properly mapped, which can result in internal or personal data appearing in AI responses.

System prompt and configuration leaks

Internal instructions, security rules, and tool definitions can often be extracted from AI applications and used for targeted attacks.

Supply chain risks

Pre-built models, plugins, and MCP tool integrations introduce dependencies that are rarely vetted.

These vulnerabilities are systematically documented in the OWASP Top 10 for LLM Applications and the MITRE ATLAS framework — the established knowledge base for attacks on AI systems. They form the technical foundation of our testing methodology.

LLM and AI pentesting: our approach

Our methodology is based on established frameworks and covers the entire attack chain.

Reconnaissance
Model fingerprinting, analysis of the RAG pipeline, mapping of available tools and interfaces.

Data layer
Access controls on RAG indices, extraction of sensitive training or context data.

Infrastructure layer
Securing model hosting, cloud ML services, and underlying Kubernetes/container environments.

Application layer
Prompt injection (direct and indirect), jailbreak attempts, manipulation of multi-agent systems.

Tool and interface layer
Abuse of function calling and MCP integrations, privilege escalation via connected systems.

Get a free consultation

Are you still unsure about commissioning AI and LLM penetration testing? Do you have specific questions about the process or the costs? Or are you concerned about potential impacts on your ongoing IT operations?

In a free consultation, we will provide all the answers you need to make an informed decision!