Obsigen · Trust / FAQ · as of 2026-10-22
For DPOs, security teams and management: where data is stored, what the model processes, which documents you receive and how to report vulnerabilities.
The wider background - shadow AI, data transfers outside the EEA, model memory and AI Act deadlines - is covered in our AI Act & GDPR guide.
We describe storage and processing separately - for each layer of the system.
data
Stored on the tenant’s dedicated infrastructure: the application, the database and files - in Managed Cloud on an OVHcloud instance in an EEA region, in Self-hosted and On-premise in your environment. The shared GRID layer has no database and does not keep task content; the browser does not connect to GRID. Tenant files, conversations and passwords written to GRID: 01.
identity
The token for every task is signed by the tenant server with its private key. GRID knows only the public key and cannot issue a token on your behalf. Forge and Runtime have one token per device, held in the OS credential store and revocable from the account panel.
models
The model receives only what the task requires. Open-weight models run at providers in the EEA; providers and regions are listed in the data flow map, shared under NDA. In the On-premise deployment model, the models run in your data center.
search
A search provider in the EEA receives the search query - not the conversation content or files.
code
Code from a task runs in a container on the tenant’s infrastructure; the network is closed except for selected package registries. Commands on your computer are classified before they run, and consent is given in a dialog on the device. On macOS and Linux, an OS sandbox is added; on Windows it is in preparation, and until then Forge asks for confirmation more often.
image
The background of a product photo is computed on the tenant server. The product itself is left unchanged, and the photo is not sent to the model provider.
What crosses each system boundary and what is stored on each side.
| boundary | what crosses it | what is stored on this side |
|---|---|---|
| User → tenant | questions, files, the entire conversation history | everything - files and conversations are stored here, in the tenant |
| Runtime / Forge → tenant | only what you choose to share from your own disk | on your side; Runtime and Forge connect only to the tenant |
| Tenant → Obsigen GRID | the content of the turn: the question and the context needed to answer it | nothing - GRID is stateless |
| GRID → models | the question and the context of the turn | nothing - Zero Data Retention written into the DPA |
| GRID → search | only the search phrase, without files or conversation history | nothing - Zero Data Retention written into the DPA, a provider in the EEA |
| GRID → web pages mentioned in the conversation | only the address of the page to be fetched | nothing of yours - the site owner sees only that the page was fetched |
| GRID → image editing | the image submitted for editing | the material is not used for training, but the provider does not declare zero retention - the feature can be disabled |
| GRID → tenant | the result - the only connection in this direction | in your tenant |
Files and conversations are stored on the tenant's dedicated infrastructure. For a model to carry out an instruction, the content of a single task is processed by GRID and the model provider - that is why we choose providers whose contracts state that they neither store it nor use it for training.
We share the full data flow map and the system architecture description under NDA; the ways to run Obsigen are described under deployment models.
Obsigen does not train models on your data. Your data is stored only in your tenant - on infrastructure dedicated to your company. This also applies to memory, the company glossary and the document index (RAG): all of them are stored only in the tenant. Memory and the glossary are built from what users say - every entry is shown as “Remembered · Undo”. The model selection table is built from our tests and synthetic data, not from your tasks.
Language model and search providers have zero retention written into their data processing agreements: the content of a turn is neither stored by them nor used for training. You approve the list of providers before the start - with the right to strike out any provider you do not want.
These are two separate commitments: a provider can refrain from training on the material and still retain it for some time (logs, buffering, moderation). That is why we state both for every provider - including image editing, where the provider does not declare zero retention.
We handle vulnerabilities and incidents under our published security policy. How we are preparing for the obligations of a software manufacturer in the EU is described on the CRA page. We also verify security with internal red-team tests.
Its own database and disk. No shared database with compartments - tenants cannot see one another.
A container without system privileges, with a read-only file system outside the working directory and limits on memory, processes and time. It disappears after a single command.
A type check, an antivirus scan and removal of photo metadata before a file reaches the conversation.
The contract sets out how data is exported and deleted at the end of the cooperation.
Obsigen Runtime connects only to your tenant, listens on no port, masks the known locations of keys and passwords before every command, and the file changes of every turn can be undone.
We show deployment files, configuration and test results once a non-disclosure agreement has been signed.
Isolation principles are described in the system architecture; work on your computer is covered under Obsigen Runtime.
The terms and conditions, data processing agreement (DPA) and KSC statement (Poland’s National Cybersecurity System Act) are part of the document package.
We share the document package with security and legal teams during deployment talks, once a non-disclosure agreement (NDA) is signed:
File quotas and memory retention are described in the DPA and the data flow map.
Obsigen does not hold its own ISO 27001 or SOC 2 certifications and does not claim them in any form. GDPR compliance of the processing is assessed by the data controller, and AI Act obligations are split between the provider of the AI system (Visera) and the company using it. We provide what your assessment can be based on: documents, the architecture description and, once a non-disclosure agreement is signed, the configuration as well: deployment files, isolation settings and test results.
In the Managed Cloud model, the OVHcloud infrastructure holds ISO/IEC 27001, 27017 and 27018 certifications; Public Cloud instances also hold HDS. Scope for individual services: see deployment models.
The information that the user is working with an AI system is permanently visible in the application interface. The status of marking images, documents and text is presented in the document package.