Privacy first security: what leaves your Mac

In one minute

To stop bad software, a tool has to see programs and settings on your Mac. It does not have to send your documents, mail or messages anywhere.
Ask where the decision is made. A tool that judges on the Mac sends a result. A tool that judges on a server first has to send the activity.
Cybee sends facts about each Mac: which protections are on, alerts with secrets removed, app and extension lists, which AI assistants run and the Mac’s name. File contents, passwords, mail and browsing history never leave the Mac.
Machine learning spots unusual behaviour by scoring it against examples. It raises the same two questions as any tool: what data does it learn from, and where does it run.
Before you install anything, ask the vendor six plain questions about your data, and keep the answers.

On this page

What does a protection tool need to see?

Programs and settings. To stop bad software, a tool has to know which programs start, what they try to do and which safety switches on the Mac are on. That is a different job from reading your proposals, your mail or a client’s files.

The two are easy to mix up. A tool that records everything can claim it needs everything. What matters is where it makes its decisions and what it sends away afterwards. The activity data a tool sends back to its maker has a name, telemetry.

What is telemetry?

What telemetry means.

Telemetry is the stream of activity data a tool collects on a device and sends back to its maker: which programs ran, which files were touched, which settings changed.

Vendors use it to spot attacks and to improve the product. How much a tool sends, and how long the vendor keeps it, differs from tool to tool, which is why the questions further down matter.

Where does the decision get made?

Picture a program on a Mac that starts changing hundreds of files in a row. Something has to decide whether that is a backup or an attack.

In one design, the Mac records the activity and ships it to the vendor’s servers. The decision happens there and the answer travels back. The vendor sees more, so it can learn from many customers at once. You pay for that in two ways. A copy of your activity now exists outside your control, so you have to ask how long it is kept and who can read it. And the tool can only act as fast as the round trip allows.

In the other design, the Mac decides. The tool watches programs on the Mac, judges there and acts there, then reports what it did. The vendor receives a result, not the raw activity. You give up some of the vendor’s view across all its customers. In return you have a much smaller pile of data to worry about.

A tool that judges on the Mac sends a result, not your work.

Neither design is dishonest by default. Both give you a test you can run in a minute: ask what leaves the Mac.

What leaves the Mac, and what never does.

Cybee uses the second design. It watches for known malware and for behaviour that has no signature yet, and acts on the Mac first. Protection, safer settings and the fixes run there too. What reaches Cybee is a list of facts. Cybee watches programs, not people.

Cybee on one Mactwo lists

what leaves the Mac

Facts about the Mac.

Which protections are on, and since when
Alerts, with secrets removed
App names and versions, browser extensions and startup programs
Which AI assistants run, and which tool servers they use
The Mac’s name, the owner’s name if set, and the macOS and Cybee versions

what never leaves the Mac

What is inside it.

The contents of your files and documents
Email and messages
Passwords, keys and tokens
Files held in quarantine
The full activity log of your AI assistants
Browsing history, the screen, keystrokes and location

The two lists for one Mac. They are the same on every Mac.

One permission needs plain words. Cybee asks you to switch on Full Disk Access in System Settings. Without it a protection tool can only see parts of the Mac, and Cybee says so in every report. With it, Cybee protects the whole Mac instead of system and application locations only. File contents still never leave the Mac.

Cybee is a Swiss company under Swiss data protection law.

That law, the revised Federal Act on Data Protection, has applied since 1 September 2023, according to the Swiss Federal Data Protection and Information Commissioner.

Machine learning: what it needs, and where it runs.

Many security vendors say their tools use machine learning to spot unusual behaviour. In plain terms, someone shows a program thousands of examples of ordinary activity and of harmful activity. The program learns to give new activity a score, and the tool acts when the score is high.

What is a learned model?

What a learned model is.

A model is the part that does the scoring. Training is the step before it: the vendor feeds the program examples until it separates the two kinds well enough.

After training, the model is just a file. It can run on the vendor’s servers or on your Mac. Two things decide the privacy cost: what went in during training, and what has to be handed to the model each time it scores something.

Those two things are where the trade offs sit.

First, what the model learned from. Examples come from somewhere: a malware lab, a public collection or the activity of earlier customers. Ask which. In December 2024 the European Data Protection Board said that whether a model trained on personal data counts as anonymous must be judged case by case. The model qualifies only if personal data can hardly be pulled out of it, directly or by questions put to it.

Second, where the scoring happens. If the model runs on the vendor’s servers, the activity it scores has to travel there. If it runs on the Mac, only a score or an alert needs to leave. That keeps more at home. It has costs too: the model has to fit on a laptop and share it with your work, and the vendor has to ship updates to it.

A learned model finds what is unusual, which is not always what is bad. A designer who opens a hundred large files on deadline day looks odd to a model that has never seen her deadline. Good vendors can tell you how they handle such false alarms.

Models are also targets. NIST’s 2025 overview of attacks on machine learning lists poisoning, where an attacker corrupts the examples a model learns from, and privacy attacks that try to pull information about the training data out of a model. ENISA, the EU’s cybersecurity agency, named data poisoning, adversarial attacks and data exfiltration as threats to machine learning systems in its 2021 report.

Cybee does not leave this decision to a model on its servers. Detection, safer settings and fixes run on the Mac, and no AI model in Cybee’s backend decides what is malicious.

Six questions to ask any security vendor.

Put these in the email you send before you install anything. A vendor that has thought about them answers in a few specific sentences. The principle behind them is old: data minimisation, collecting only what the purpose needs, is one of the named principles of the EU’s General Data Protection Regulation.

before you installsix questions

What exactly leaves my Mac? Ask for the list, item by item.
Are file contents, email, messages, passwords or browsing history on that list?
Is the decision made on the Mac or on your servers?
If you use a learned model, what was it trained on, and does it learn from my activity?
How long do you keep what you receive, and who at your company can read it?
What happens to my Macs’ protection if your service stops?

Six questions about data. Nobody ticks them. Write down who answered and when.

You can hold Cybee to the first two. The lists above are the answer, and the security page sets them out with the services you choose to connect.