The detection engine
What Bombus is, how the engine recognises sensitive data, where detection runs and what our own test says about its accuracy.
All detection in BeeSensible runs on Bombus, BeeSensible's detection engine. Every organisation uses the same engine, whether the text is in the browser, in the desktop app or in a document. This article explains how Bombus reaches a result, where it runs and what our own test says about its accuracy.
How Bombus recognises sensitive data
Bombus, currently version 2.1, combines two ways of recognising data. For data with a fixed form, such as bank account numbers, national ID numbers, card numbers and email addresses, it uses fixed rules. For everything that depends on context, such as names, addresses and medical or other sensitive details, it uses small, fast AI models. Bombus merges the results of both into a single result, so overlapping matches become one highlight.
Bombus recognises 65 data types in Dutch and English, grouped into 11 categories: person; location and address; contact; identifiers; organisation; medical; financial; technical; demographic; travel; and time. You decide which of these data types are highlighted in your organisation, and how strictly, through detection profiles. Terms that only mean something within your organisation, such as project names or the names of internal systems, can be added as custom terms.
Where Bombus runs
Bombus runs on BeeSensible's own servers at European providers: the application at Scaleway in Amsterdam and the models on a dedicated server at Hetzner in Germany. Both are European companies, so BeeSensible does not use the European region of an American cloud provider. Both providers are ISO 27001 certified.
The extension and the desktop app send the text to these servers. Bombus analyses it in memory and discards it straight afterwards, so BeeSensible keeps no content. No external AI service is involved. What BeeSensible does keep, for each highlight, is the data type, the level, the app and the time, with no user ID and no text.
What our own test says
On our own test set, Bombus achieves an F1 score of about 96% across all 65 data types, in Dutch and English. The F1 score combines two things: how much of the sensitive data the detection finds, and how much of what it highlights is actually sensitive. This is an internal test on a frozen test set that we did not use to tune the engine, so it is not an independent measurement and not a certification. The figure describes what the detection finds and says nothing about how many data breaches it prevents.
Last updated:
