Protect company data, your website and your AI from misuse.
We help you limit leaks of sensitive data when people use AI, spam in your forms and attempts to misuse your chatbots and applications. We deploy the protection for your company, or connect it into the product you provide to your own customers.
Where do you need protection?
Employees use AI
Contracts, e-mails and client data end up in a chat easily. Help the team work with AI while sensitive information stays protected.
One Corpilus protection. You choose where you need it connected.
See what changes once the protection is on.
These are a few examples from the protection, not a complete list of what it can do.
An employee pastes a paragraph from a contract — the client's name and the IBAN — into a public chat (ChatGPT, Gemini, Claude) and attaches a document to it.
The protection recognises the sensitive values in the chat window and, per the setting, replaces them before the text is sent.
You see the original text and what the model actually receives, side by side.
The check also covers the text of an attached document — in the supported browser chats (ChatGPT, Claude, Gemini). Images and unsupported file types stay unchecked and the protection says so; in the other connection modes what is checked is the text the application hands over.
A bot tries other people's passwords at your sign-in page, and an agent downloads the content of your site page after page.
The protection records every sign-in attempt as its own event — a failed one apart from a completed one, and never with the password or the e-mail address — checks a visitor that calls itself a known crawler against that vendor's published addresses, and recognises catalogue pages being downloaded over and over from one device.
In the overview you see who tried what and why it was flagged. A sign-in is refused by your own credential check, never by the protection.
A sign-in attempt is recorded only when your server hands it over. What a visitor calls itself proves nothing on its own, and where the address cannot be checked the finding says so rather than call a genuine crawler an impostor.
A document the chatbot reads carries an instruction aimed at the model, and the agent is about to make a tool call it is not allowed to make.
The input check marks the instruction before the content is processed, and the call's arguments are compared with the tool's declared schema in the connected gateway.
You see the finding together with its reason; what follows is governed by the channel's setting — from a notice to stopping the request.
This applies to the tools that go through the connected gateway. It does not automatically cover every tool of an agent.
Fake enquiries keep arriving through the contact form — the same text over and over, and submissions that fill in a field no visitor ever sees.
The protection evaluates the submission from the signals of the connected channel and writes the finding, with its reason, into the event overview.
You see why the submission was flagged; stopping it is done by the connected application, once it verifies the result of the check on its server.
The script on the page adds the hidden field and reports the signals. It does not stop a submission by itself.
Someone pushes stolen cards through the order, one attempt after another, to find out which of them still works (carding).
From the payment flow the protection receives only a fingerprint of the card, the opening digits of its range and the last four — never the whole number — and out of those it judges repeated attempts with the same card, many different cards from one range, an unusual share from a single issuing country and the shape of card testing from one device.
The payment flow gets a verdict back — from a notice through an extra check to a refusal — and applies it. A pattern other shops have already seen supports a finding; on its own it never leads to a refusal.
A payment is judged only when your payment flow hands the attempt over. The values here — card •••• 0000, range 000000 — are fictitious; a whole card number is refused at the interface.
A file arrives through a form or into the application — a document with a macro, a PDF that runs something the moment it is opened, or an image that is really a script.
The check runs in two steps. First the structure of the file: whether the content matches the declared type, whether the type and the size are allowed, whether a document carries a macro, a PDF an action that runs on opening, an SVG a script. Only a file that passes goes on to the signature check by the antivirus engines.
You see the verdict together with the finding. A file that no engine managed to check is reported as unchecked — never as clean.
Whether the upload is stopped or set aside is decided by the connected flow and its setting. The example file offer.pdf is fictitious.
Sample scenarios with fictitious data. The scope of the checks and the reaction depend on how the protection is connected and configured.
Each tab checks one connected channel — the browser, the form handler or your application's interface — and nothing outside it. Integration options
We start from what you need to protect.
We pick one concrete situation.
Using AI, a website, or your own application.
We start the measurement and collect data.
We first turn the protection on in monitoring mode, so you can see what is really happening in the connected channels.
We agree how it connects.
The access needed, the responsibilities and the way the result is verified.
We set the reactions.
According to the findings and according to how you operate.
We go through the deployment and operating options according to where you connect the protection. Deployment options
There are people behind the product you can reach by name
A company doesn't buy a security layer from an anonymous vendor.
A support team works with them on deployment, operations and support.
Do you already have a WAF or another protection?
Corpilus does not replace it and does not switch it off — it joins it as a further layer. It adds checks where what matters is the content of a request and the data heading for AI. What it actually adds in your setup depends on the solution you run and on how that solution is configured.
Compare it with the protection you already runWe will show you the protection on a situation from your company.
Write to us where you use AI or what you need to protect. We will go through a suitable way to connect it and propose the next step.
The agent sends a request. The protection assesses it and marks a finding. The request is stopped and the reason is in the event overview. A sample run, not a measurement.
corpsys.desk — the company environment for documents, data and automation
- Company AI over documents
Ask in your own words; the answer comes from your own files.
- Secure data
Company data stays in an environment you control.
- Automation
Recurring work runs as agreed, repeatable steps.
- Service desk and 24/7 monitoring
Requests in one place; website and operations watched around the clock.