01
Pentesting
We try the same tricks on your AI that a real attacker would. You get a clear list of weak spots, advice on how to fix them, and material for NIS2 and DORA.
What we test for
Prompt injection
Instructions hidden in user input, documents or web pages that take over your model or make it ignore its own rules.
Data leakage
Your system prompt, customer data or secrets ending up in an answer.
Unsafe actions
Agents calling tools, running commands or sending data they shouldn't.
Hallucinations
The model confidently making up facts, policies or prices and presenting them to your users as true.
02
AI firewall
It sits between your users and your AI and watches every request and answer. It works with any model, even one we can't see inside.
- Blocks attacks before they do any harm.
- Monitors everything. If something happens, you know why.
- Watches what agents do, not just what they say.
- Works with ChatGPT, Claude, Gemini and other models.
“”▍
03
Open-source AI under control
With an open-source model we look right inside. We recognise a dangerous message by how it lights up inside the model and stop it before the model answers. We can then fine-tune the model for you, so it withstands the attacks your company actually faces.
No added latency.
The model does these calculations anyway. We simply read what it has already computed, so your AI doesn't slow down at all.
“How do I bake sourdough bread?”
Message passing through the model…
The research we build on
Arditi et al., 2024
Refusal in Language Models Is Mediated by a Single Direction
Whether a model refuses a dangerous request is decided by a single direction inside the model. That's what we watch. →
Cunningham et al. (Anthropic), 2026
Constitutional Classifiers++
Checking the model's activations directly costs a fraction of the compute of a separate classifier and holds up in production. →
Zou et al., 2024
Improving Alignment and Robustness with Circuit Breakers
A model's harmful “thoughts” can be interrupted inside the model, before they turn into an answer. →
Zou et al., 2023
Universal and Transferable Adversarial Attacks on Aligned Language Models
Why testing matters: automatically found attacks get around model safeguards and work on commercial chatbots too. →