Zum Inhalt springen

Schwerpunkt: Lokale KI und Pseudomizer. Zusammen. Mehr

Branchenpartner gesucht. Partner werden

Service

The risk is not in the model.

An assistant that reads incoming post also reads what someone has written into it to control it — in an attachment, in white text, in the metadata. The system does not reliably distinguish between your question and what is in the document.

Stack of paper behind a pane of glass

Checking permissions

Separating third-party text

Reporting route before the emergency

What it is about

Three ways in which it goes wrong.

Third-party instructions

Incoming documents can contain text that is not addressed to you but to your system. ⛔ This is not a legal category but a property of this class of system — and the reason why an assistant may read but must never send on its own.

Excessive permissions

Connections to mailboxes, file stores and databases are almost always set up with write access because it is easier. Whatever can write can also delete — and nobody plans for that in advance.

No route for an emergency

If you suspect a data leak, every hour counts: the report to the supervisory authority is due within 72 hours. Anyone who only starts looking for the route then loses the first day.

How it works

Four steps that belong together.

01

Stocktake of access rights

Which tool reaches which file store, with what rights. It usually turns out: more than expected.

02

Cutting back permissions

Read-only instead of write access wherever possible. Separate data stores where things must not come together.

03

Securing the inbox

Third-party text is kept separate from the task. And the hard rule: no tool with its own sending route.

04

Log and reporting route

Who asked what, what came back. Plus a reporting route that is written down before the emergency and not devised in the middle of it.

What is there afterwards

What is secured afterwards.

  • Interfaces are connected read-only wherever write permissions are not needed.
  • Data stores that must stay separate are separate — not by agreement, but by permissions.
  • Incoming third-party documents do not run unfiltered into the task.
  • There is a log: question, answer, source reference, time.
  • A reporting route with a name and a deadline, which does not have to be looked for when it matters.

And what it does not

This is here because everything else would be worthless without it.

Anyone who promises everything is switched off at the first objection. That is why we say beforehand where the limit lies.

It does not make you unassailable

Security is not a state you can buy. We reduce attack surfaces and make incidents noticeable — we cannot rule them out, and anyone who promises that is not telling the truth.

It does not replace your IT security

Firewall, backups and rights management remain what they are. We take care of what is newly added by AI tools.

It is not continuous operation

We set up and hand over. Ongoing monitoring would be something different and would have to be agreed separately.

A look inside

The airlock from the inside.

A two-minute flight through the operational view of our airlock. Every body is a programme, every line a route along which data travels. You see where secrets are replaced, what stays local and what the cloud never receives.

Status: version 0.1 alpha, so far tested only with fictitious examples. The view is a snapshot from 05.10.2026, names are neutralised. Not legal advice. Load file

Frequently asked questions

The questions that come first.

What exactly is a 'third-party instruction'?

A sentence in an incoming document that is addressed not to the reader but to the system — for instance an instruction to ignore earlier instructions or to forward content. It can be invisible: white text, metadata, footnote.

Is it enough to tell employees to be careful?

No. The text is often not visible to people. ⛔ Attention helps against obvious cases, not against those written for machines.

Do we need new software for this?

Mostly not. Most of it is setup: rights, separation, logging, reporting route. That costs care, not licences.

What about screen recordings?

Some tools can read the screen. During a meeting in progress or with a second system open, that is a problem. The function should be switchable off and blocked for confidential windows.

And what if the provider itself is attacked?

Since 11 September 2026, manufacturers must report actively exploited vulnerabilities to the authorities within 24 hours and inform their users. ⭐ Have the contract guarantee that you will be informed — not only from the press.

How we start

First your biggest annoyance. Then the rest.

Bring us your biggest annoyance.

One conversation, one real task from your daily work. We use it to show what AI and automation can do for you.