AI features that do not leak.
If you have wired a language model to something that matters, it is a new attack surface. I test it as an attacker and build the defence that has to hold afterwards.
WHAT IS INCLUDED
- Prompt injection testing against your own features, not against an example
- Jailbreak and role-switch resistance: can the model be talked out of its rules?
- System prompt extraction: what can be pulled out of your instructions
- A review of what the model can reach, and what it should not be able to
- Guardrails in the code, not only in the prompt — a boundary has to be enforced where it cannot be argued with
- Logging of rejected attempts, so you can see it happening
- Written report with findings, severity and concrete remedies
WHO IT SUITS
For anyone who has wired a language model to something that matters. A chatbot that can read customer data, an agent that can write to a database, an assistant that can send e-mail in the company’s name.
HOW IT GOES
- ScopingFree. What is connected to what, which data the model can reach, and what happens in the worst case.
- AttackI run the vectors against your own setup and document what gets through.
- DefenceGuardrails go where they cannot be talked away — in the code around the model, not in the instructions to it.
- Report and retestYou get the findings in writing, and I run the same attacks again once the remedies are in place.
PRICE
Security work is priced by scope. The scoping call is free, and it is what settles the scope.
See the pricing →THE PROOF
The console on the front page is not an illustration — it runs four real attack vectors against a language model, live, and shows what an unsecured system answers next to what a secured one does instead. My bachelor thesis is about exactly this, commissioned by Telenor, tested against Gemini, ChatGPT, Copilot and LLaMA 3.1.