Security & compliance

Sovereign AI security: your data protected end to end

Configure and verify security controls on the actual deployment: hosting, encryption, authentication, document permissions, retention and monitoring. They reduce risk without eliminating it.

Our security controls

  • EU hosting

    EU location is separate from jurisdiction: assess providers, subprocessors, support access and key control.

  • Daily wipe

    Wiping GPU instances requires a verifiable procedure and hosting commitments. Restarting alone does not prove data erasure; storage, logs and backups are separate.

  • Encryption & VPN tunnel

    Encrypted traffic: users reach the portal over HTTPS, and the GPU servers join a WireGuard VPN tunnel that links them to the portal.

  • SSO authentication

    Login through your existing directory (SSO): centralised access management, no extra password, immediate revocation.

  • Control external flows

    No data is sent to a third-party AI provider or reused to train a model. The model runs on a GPU server rented in the European Union; outbound technical flows (including weights and images downloaded at startup, and email reports) must be inventoried for each deployment.

  • Licences and auditability

    Components, versions and licence conditions must be documented for each deployment. Access to source code does not by itself guarantee unrestricted use or exit.

  • Monitoring & backups

    Prometheus observability with GPU and inference probes, metrics collected on the platform's GPU servers. Backups and alerting policy defined with you during scoping.

  • Per-organisation isolation

    Your knowledge bases and histories stay within your perimeter, partitioned, never pooled with other clients.

Architecture & operations

"Operated for you" is not a sales line, it is a software layer. Here is what it actually does, and what your IT department can verify.

  • Self-hosted monitoring

    Prometheus, a system exporter and the NVIDIA DCGM exporter, with dedicated probes on the inference engine, the embedding service and GPU load. Prometheus metrics stay on the platform's servers: they are not sent to any third-party observability service.

  • Scheduled shutdown, automated

    Deleting the model's GPU server in the evening and recreating it in the morning is driven by the deployment layer, not by a manual step. That is what makes it practical to pay for the model's compute during working hours only.

  • Interchangeable host

    Provisioning, start and stop all go through a common interface that is independent of the GPU provider. Only one host is connected to it today: adding another means writing its connector, not rewriting the rest of the deployment. This is reversibility applied to infrastructure.

  • Server-to-server tunnel

    The WireGuard tunnel is installed by script on the GPU servers, which join it at every start to be linked to the portal. The portal itself is reachable by users over the internet through HTTPS, behind authentication.

  • Separate components

    Authentication, database, embedding service, inference engine and model management are distinct components. Each can be audited, updated or replaced without touching the others.

  • Reproducible deployment

    The installation of the GPU servers is described in scripts, not in a manual runbook, and the model's server is rebuilt by these scripts every morning, which makes the infrastructure verifiable and transferable.

Regulatory compliance

  • GDPR (GDPR)

    We provide the documentation for your DPO: support for the data protection impact assessment (DPIA), legal basis, minimisation, and an erasure procedure, including the vectors held in a RAG system.

  • AI Act

    We help you place your use case within the risk tiers of the European regulation and prepare the obligations and deadlines, including the fundamental rights impact assessment that public bodies must carry out before deploying a high-risk system.

    Assess my AI Act risk

Frameworks & hosting

Depending on how sensitive your data is, we design the hosting to meet the relevant security frameworks. Architecture and hosting provider are adapted to your compliance requirement.

  • SecNumCloud (ANSSI)

    The French qualification framework for trusted cloud offerings, for the most sensitive data.

  • HDS : health data

    Certified Health Data Hosting, required when a provider hosts health data on behalf of those who produce or collect it, or of the patient.

  • ISO/IEC 27001

    Information security management: governance, risk management and continuous improvement.

The hosting provider and level of qualification are defined together during scoping, based on the nature of your data.

Official sources

A specific security requirement?

Let's talk about your regulatory constraints and the architecture that fits your data.

Components, licences and operations

Open WebUI has version-specific branding terms from v0.6.6. Check the version or fork, licence and permissions for each installation before characterising the stack. A contractual inventory must identify component, version, licence and restrictions.

Open WebUI: licence

Specify service hours, startup time, fallback capacity, support, backups, recovery time (RTO) and acceptable data loss (RPO) in the contract. No numerical recovery target or 24/7 availability is presumed. HDS, SecNumCloud and ISO 27001 apply to the certified offer, activities, validity and scope, not automatically to all of PixelRaise.

Flows and retention

Users → application and access control → RAG storage → GPU computation → response. Application and GPU → logs/observability; application and storage → backups. Each layer needs access and retention policies. Wiping a GPU does not purge the index, history, logs or backups. Document injection, corpus extraction, administrator access and human error still require controls and testing.