Product / GDPR & Compliance
Compliance by Design — Not by Contract
The most fundamental compliance property of this solution is architectural: No personal data, no proprietary information, and no confidential business content ever leaves your own infrastructure.
Data Sovereignty Through Architecture
No Contract Can Replace What Architecture Guarantees
There is no data processing agreement with a cloud AI provider, no cross-border data transfer, and no third-party sub-processor for AI computations. This is compliance by design — not by contract.
Cloud AI (Status Quo)
- Data processed on external servers
- GDPR risk through third-country transfers
- Unpredictable per-token costs
- Vendor lock-in, model changes
- Rate limits, internet dependency
- Data retention contract-dependent
contboxx Vault (Local)
- All data stays on-premises
- Full GDPR compliance by design
- Fixed investment, no token costs
- Full control, model versioning
- Dedicated capacity, no limits
- Data lifecycle under your own DPO
GDPR Compliance Matrix
Checked Article by Article
Complete mapping of GDPR requirements and the EU AI Act to the contboxx Vault architecture.
| Category | Requirement / Measure | Status |
|---|---|---|
| Art. 5 | Personal data is processed exclusively on-premises, no external transfer | Fulfilled |
| Art. 5 | Data minimization: Only role-authorized documents are queried | Fulfilled |
| Art. 6 | Legal basis: Legitimate interest / contract, no consent required for internal AI tools | Fulfilled |
| Art. 13/14 | Employees are informed via privacy notice that AI processes internal queries | Configuration |
| Art. 17 | Right to erasure: Deleting a document from Atlas removes it from all AI context — no cloud retention | Fulfilled |
| Art. 22 | No fully automated decisions affecting individuals — AI is advisory tool only | Fulfilled |
| Art. 25 | Privacy by design: No external sub-processor, local architecture enforces data protection | Fulfilled |
| Art. 28 | No data processing agreement with AI provider required (no external processor) | Fulfilled |
| Art. 32 | Technical security: mTLS between Atlas and LLM API, JWT authentication, encrypted storage | Configuration |
| Art. 35 | DPIA recommended for AI-assisted HR/compliance use cases; simpler scope than cloud AI | Configuration |
| Art. 44–49 | No third-country transfers whatsoever (EU standard clauses not required) | Fulfilled |
| EU AI Act Art. 9 | Risk management: AI as internal knowledge tool is low-risk category | Fulfilled |
| EU AI Act Art. 13 | Transparency: Atlas interface labels AI-generated content | Configuration |
"Fulfilled" = architecturally guaranteed. "Configuration" = implemented project-specifically during installation.
Technical Security Measures
Enterprise Security from Day 1
The following security controls are part of every standard installation:
Network Isolation
DGX Spark Cluster and contboxx Atlas server reside in an isolated VLAN without direct internet access. AI API calls are made exclusively over the internal network.
Mutual TLS (mTLS)
All API communication between Atlas and the inference server is encrypted with certificate-based mutual authentication.
JWT-Based Authorization
Every API request contains a signed token with the Atlas identity and role of the requesting user — for seamless auditing of all AI queries.
Audit Logging
All LLM requests are logged (query hash, user ID, model tier, timestamp, response latency) — without storing the full query text. Balance between traceability and data protection.
Role Isolation
The inference layer ensures that the document context passed to the AI contains only documents for which the requesting user has Atlas read permissions.
Encryption at Rest
DGX Spark NVMe storage is encrypted. Model weights and document indices are stored on encrypted volumes.
No Training on Customer Data
The deployed open-weight models are used exclusively for inference. No user queries are used for fine-tuning unless explicitly initiated by the organization.
Patch Management
NVIDIA DGX OS (Ubuntu-based) receives security updates. The AI stack (SGLang, CUDA libraries) is maintained on a defined patch cycle.
Physical Security
Hardware resides in the organization's own data center under existing physical access controls.
Residual Risks & Mitigations
Transparent Handling of Residual Risks
| Residual Risk | Probability | Mitigation |
|---|---|---|
| Model hallucination in document summaries | Low–Medium | RAG grounding, always with source references; human review for critical documents |
| Unauthorized internal access to AI audit logs | Low | Restrict log access to DPO and security team |
| Model weight exfiltration via compromised node | Very low | Encrypted storage, no external NIC in compute VLAN |
| Supply chain risk with open-weight model | Low | Checksum verification on model download; air-gap after installation |
Compliance questions? We have answers.
Talk to our team about GDPR, EU AI Act, and your specific compliance requirements.
Schedule a Conversation