Security: eight controls on the path of every request
Security is not delegated to the AI model. Every request passes fixed checkpoints before it reaches a model and before any action is executed in an application.
The security controls in detail
Each control expands to a short description and a diagram.
Authentication and authorisationEvery action is performed on behalf of a named user.
Users sign in with the corporate account and multi-factor authentication (MFA). The AI has no permissions of its own: what it may read and do in an application is determined by the permissions of the requesting user. Where a user has no access to a folder, the AI does not use it for that user's answer either.
Data classification in three zonesData is segregated by sensitivity before the AI processes it.
Every connected system is assigned to exactly one of three zones. The zone determines which AI model may process the data, and under which conditions. Confidential content remains in a separate on-premise system.
PseudonymisationNames remain on-premise: the AI model sees pseudonyms.
Before a request leaves the corporate network, personal data is replaced by neutral pseudonyms. The mapping table that resolves each pseudonym never leaves the corporate network.
A single encrypted egress pathOne controlled path to the AI model, initiated from the inside only.
Between the corporate network and an AI model in the cloud there is exactly one path: a security proxy between two firewalls, then an encrypted tunnel that is initiated from the inside only.
Data residencyThe organisation determines where its data is processed.
The AI can run in the EU cloud, on-premise or in a data centre of choice. In the EU cloud, requests and answers are not shared with the model provider and are not used to train AI models.
Two checkpoints: AI gateway and MCP gatewayThe AI model proposes; it does not act autonomously.
Every action passes a checkpoint, and a checkpoint is conventional software with fixed rules. The AI gateway sits in front of every AI model and inspects input and output. The MCP gateway sits in front of the business applications and permits only actions from an allow-list.
Human approval for critical actionsCritical actions require the approval of a person.
Which actions may be performed without approval is defined in advance. High-impact actions are presented to a person first and are executed only after approval.
Audit logging, alerting and emergency shutdownEvery step is logged and monitored, and the system can be shut down.
For every step, the audit log records who made the request, which tool was used and what the checkpoint decided. Anomalous behaviour raises an alert. The system can be shut down at three levels: a single connector, a whole zone, or the entire system.
Aligned with the guidance of the BSI
The design follows the published guidance of the BSI, the German Federal Office for Information Security. The cloud provider that runs the AI model holds a C5 attestation of the BSI.
The figures and documents behind these pages: Sources.
Your AI. Your data. In the EU.
Discuss a use case with us
20 minutes, without obligation: the situation is described, and we say what AI can realistically do, and what it cannot.
Open the form
- contact@ingnetworks.com
- Phone
- +49 1515 3851511
- Address
- Ingenious Networks GmbH
Eisenacher Straße 6
10777 Berlin, Germany