AI penetration testing
Structured testing of your LLM application or agent against the OWASP Top 10 for LLM — manually and with our own attack tooling. Report with prioritised findings and concrete fixes.
// ai security
Can your AI be manipulated?
Chatbots, copilots and AI agents are the newest attack surface in your business. We test them the way an attacker would — prompt injection, data leaks and tool misuse — and build the defences that hold.
Book an AI security testWhy AI is a new attack surface
A language model does not distinguish between instructions and data. Everything it reads — a customer message, an attached PDF, a web page, an email — can contain hidden instructions that make it leak data, ignore its rules or act on the attacker's behalf. Prompt injection has been number one on the OWASP Top 10 for LLM applications since the list was created. And with AI agents that can send emails, look things up in systems and take actions, the consequence is no longer an embarrassing answer but a real breach.
What we test for
Prompt injection
Direct attacks in the conversation and indirect attacks hidden in documents, emails, web pages and data the model retrieves.
Data leaks
Extraction of system prompts, internal documents from RAG, other users' data and secrets in the context.
Agent misuse
Can an agent be tricked into sending, deleting, buying or changing something it shouldn't? Excessive agency and tool misuse.
Jailbreaks & unsafe output
Bypassing rules and policies — and output that ends up as XSS, SQL injection or commands in your systems.
Data poisoning
Manipulation of the documents and sources your AI builds its answers on.
Access & permissions
Do the model and its tools have more permissions than the user talking to it?
What we offer
Architecture and code review
We review how the AI is wired to your data and systems: permissions, tools, RAG pipelines and what happens with the output.
Guardrails & hardening
Input and output filtering, tool isolation, least privilege for agents, logging and monitoring of AI traffic.
Secure AI integration
If you want to build AI into the business, we build it securely from the start — and run it afterwards.
FileSanitizer for AI pipelines
Files that will be read by your AI or your people are sanitised first (CDR), so hidden instructions and malware never reach the model.
Ongoing re-testing
Models and prompts change. We test again when you update, so security does not erode.
How it works
01
Scope
We map your AI solution: model, tools, data sources and who can talk to it.
02
Test
We attack in a controlled way — in the conversation, via documents and web pages, and through the tools the agent has access to.
03
Report
You get the findings prioritised by real risk, with reproducible examples and concrete fixes.
04
Fix & re-test
We help close the holes and test again, so you know it works.
// questions & answers
Frequently asked questions
Have a question that isn't covered here? Email us at info@edgetek.dk
01 What is prompt injection?
Prompt injection is when an attacker gets a language model to follow their instructions instead of yours — either directly in the conversation or indirectly via content the model reads, such as a PDF, an email or a web page. It can lead to data leaks, bypassed rules or unwanted actions.
02 We just use ChatGPT or Copilot — is this relevant for us?
Yes, if the AI has access to your data, documents or systems — for example via plugins, RAG or agents. The risk is not in the model alone, but in what it is connected to and what it is allowed to do.
03 How does an AI pentest differ from a regular penetration test?
A regular pentest finds vulnerabilities in code and infrastructure. An AI pentest also tests the model's behaviour itself: can it be manipulated, leak data or misuse its tools? We usually do both, because AI solutions have classic vulnerabilities too.
04 What do you need from us?
Access to the solution as a regular user, a description of what the AI has access to, and a point of contact. We test within an agreed window and without loading production.
Book an AI security test
Tell us briefly what your AI solution does and what it has access to. We reply within one business day.