trust center · as of 2026-04-20
Our full status on security, privacy and governance. Not as marketing — as accountability.
section b
For governments, sovereignty is not a preference but an obligation. EU AI Act, GDPR, NIS2, BIO — every regulation imposes requirements on how you handle data, models and decision-making. We have not translated those requirements into a checklist page. We have built them into the architecture.
section c
Four classes. Per class: who may see it, which AI model may touch it, where it may run.
L2 — Geheim
Persoonsgegevens, klachtdossiers, medische info. Draait op een zelf-gehost open model (Llama, Qwen, Mistral) via Ollama, binnen klant-infrastructuur of een MurmurWorks OnPremise-omgeving. Data verlaat de klantomgeving fysiek niet — ook niet voor embeddings of indexering. Strikt need-to-know toegang, altijd gelogd.
Eén platform. Vier veiligheidsklassen. De classificatie bepaalt automatisch welk AI-model de data mag raken. Geen instelling — architectuur.
section d
Security Model Selection, EU AI Act, GDPR, human oversight — click through for full detail.
Murmur | Works chooses the AI model per task and per data sensitivity. A public news article may hit a frontier model for maximum quality; a complaint file with special-category personal data physically cannot reach a commercial cloud. The choice is not a setting — it is architecture.
How it worksHuman oversight, risk management, technical documentation, audit log — the Act's obligations are built in, not bolted on. Working on a high-risk application? We activate the heavier set (conformity assessment, EU database registration).
Our Act positioningYour data stays on Dutch or EU soil. We are processor, you are controller — with a standard data processing agreement per art. 28. Extra-sensitive data is stored strictly separated, with its own security. All data is encrypted — at rest and in transit. When a citizen requests erasure, we actually erase everything: raw data, derived analyses, copies.
Request DPIA templateEvery outgoing action requires human approval. Every AI call is logged — which document, which model, which prompt, which response, which classification. Stored immutably, reported per quarter in anonymised form.
Sample transparency reportsection e
| Standard / framework | Status | Target |
|---|---|---|
| GDPR conformity | Secured in architecture + DPA | Ongoing |
| EU AI Act | Conforming | Phased: transparency Aug 2026, high-risk Dec 2027 |
| ISO 27001 | In preparation | Q4 2026 |
| NEN 7510 (healthcare variant) | Under consideration | Depending on market demand |
| BIO (Dutch Gov Baseline) | Gap analysis planned | Q3 2026 |
| NIS2 | Depending on client categorisation | Research in 2026 |
| Pen-test (external) | Annually from 2026 Q3 | Ongoing |
| SOC 2 Type II | Not planned | EU market, ISO 27001 prevails |
section f
What we don't have (yet):
section g
Includable in contracts.
100%
of data above L1 never forwarded to commercial cloud providers
< 30 days
processing GDPR access or erasure requests
100%
of outgoing actions require human approval
1 business day
full audit-trail access for the client
0%
client training data leaves their environment for model training
section h
Specific security or privacy questions from your counsel, DPO or CISO? Request our security dossier — then we schedule a substantive conversation, not a sales call.
Request security dossier