Privacy that is enforced, not promised

Security & privacy

The interesting question is not whether a vendor says your data is safe. It is what the code does when nobody is looking — which parameters, which algorithms, which failure modes. This page is the detail; the source is one click further.

The controls

Four layers, and what each one actually does

Zero-knowledge encryption

Credentials, sealed separately

Audit that does not become a second leak

Compliance you can evidence

Identity, brought in from your IdP

Two things this page deliberately does not cover

Where to read further

Outbound data-loss prevention has its own page, because the interesting part is what the detector does to a prompt rather than how it is configured — see Privacy Shield for the tokenise-and-restore walkthrough. Where your data physically sits, which model provider sees it, and how that is measured has its own page too. Both are linked below rather than summarised here, so there is one place per subject that stays correct.

Where things stand

Honest about the boundaries

The security posture in numbers

FAQ

Security questions worth a straight answer

Can Bee Flow staff read our documents?

On a self-hosted install there is no Bee Flow staff involved — you hold the infrastructure. On the hosted version, your chat and notebook conversations are encrypted under a key derived from your own password with Argon2id, and there is no support backdoor that would let us unwrap them. Knowledge-base documents and meeting transcripts are stored server-readable, because search and diarisation have to read them — we would rather name the boundary than let the strongest claim cover everything.

What happens when someone forgets their password?

They use their recovery key, which wraps the same data key independently. Without either one the encrypted data cannot be recovered — that is the direct consequence of there being no backdoor, and it is worth telling your users before they need it.

Does the PII detection send our text to a third party?

No. Detection runs in a container on your own hardware using an Apache-2.0 licensed model on CPU — 21 categories, with the Dutch BSN validated by its checksum rather than guessed. An external detection service can be configured if you want one, but nothing on the default path leaves your network. If you install the detector and it later becomes unreachable, the request fails closed rather than being sent unchecked. Privacy Shield has the full walkthrough.

How do we prove any of this to an auditor?

The compliance hub runs 15 GDPR checks and 6 EU AI Act checks against the live configuration and shows what passes, what fails and why. Combined with the guardrail event log and the automation run history, that is an evidence trail rather than a policy statement.

Do not take the security page on trust

Run it yourself and watch the network, or read the encryption and DLP code directly. Both are faster than a vendor questionnaire.