This page is machine-translated from French. Read the French original.
Security and confidentiality
The principle
Conversations and data remain local only if the language model and embedding engine are local. In a fully local configuration, documents, indexes, conversations, models, and the database remain on your infrastructure.
Each content output results from your configuration: connecting an external provider, a hosted database, a search engine, an agent skill, or an MCP server. WivenLLM does not block any of these outputs by default. It logs them, and it is the strict sovereign mode that blocks them.
Other connections do not carry content, such as downloading a template on its first use, or checking for updates by the desktop application every six hours, which strict mode cuts off.
The Settings → Privacy and Data screen (link at the bottom of the menu, administrators only) summarizes the model provider, embedding engine, and vector database configured for the instance, as well as the telemetry status. It does not show the space-specific model, transcription, or skills. To see the actual output: Wiven Console → Sovereign Mode.
Where the data goes, component by component
| Component | In local configuration | In external configuration |
|---|---|---|
| Language model | Nothing comes out | Messages + documentary context |
| Embeddings engine | Nothing comes out | Text of documents and questions |
| Vector base | Nothing comes out | Vectors and metadata |
| Audio transcript | Nothing comes out | The audio file |
| Text-to-speech | Nothing comes out | The text of the answers |
| Image generation | Nothing comes out | The description |
| Web search (agent) | Not applicable | The request |
| URL import | Not applicable | Request to the requested address |
| Messaging and calendar (agent) | Not applicable | Reading and writing at Microsoft or Google, using the person's account |
| MCP Servers | According to the server | What the agent transmits to them |
| AgentHub Connect (if connected) | Not applicable | Model, tokens, duration, data class, number of documents; never the content |
The most frequently overlooked aspect is the embedding engine. A local model combined with an external embedding engine exposes the entire corpus to that provider for indexing. The choice of embedding engine deserves as much attention as the choice of model. The privacy shield does not apply to this data stream.
The protections and their default state
Most protections are in place but need to be activated. They are configured in the Wiven Console: a button next to the model selector, visible to administrators and managers, or to the single user in single-user mode. Most tabs are reserved for administrators.
| Protection | Default state | Namely |
|---|---|---|
| PII Shield | Disabled | Replaces personal names, social security numbers, IBANs, IDE numbers, QR code references, cards, email addresses, and phone numbers with pseudonyms in the data sent to the language model, then restores the response. Agent tools receive the real values. Embeddings and other external services are not covered. Detection is rule-based and may miss data. |
| Sovereign Mode | Monitoring: logs, does not block anything | Strict blocks outbound traffic outside the company network and outside the whitelist, including for the document collector. It applies to outbound connections from platform services. Behind a proxy, only the proxy is visible. |
| Content control | Disabled | Rules based on patterns, without understanding the meaning. → Content control |
| Unsourced claims | Disabled | Adds a «To be checked» note to document-based answers that do not contain figures or sentences in the passages. Answer Check tab |
| Approved models | Disabled | File fingerprinting required for embedded engine models. Does not affect other providers |
| Walls and rights by document | No rules | Applies to all roles, but not to API keys. → Users, roles and organizations |
| Two-factor authentication | Not required for any role | → Users, roles and organizations |
| AI Transparency | Activated | Marks generated files (PDF, Office, PNG, JPEG), not text or video. Warns widget visitors and Telegram users that they are communicating with an AI. |
| Audit log | Active | → Traceability, below |
Prompt injection
Like any model-based language system, WivenLLM cannot guarantee that an instruction hidden within content will be ignored: a document, web page, or email may contain instructions that the model will follow. The safeguards are limited: instructions given to the model, human approval of the agent's actions, and, for reading web pages, rejection of internal addresses chosen by the model.
Sending, replying to, and deleting messages always require approval, even if the agent is authorized to act independently; execution without approval (scheduled task, API) is rejected, and a maximum of five messages are allowed per execution. Do not grant an agent access to both untrusted content and sensitive skills.
Protect the instance
Installation secrets
Installation secrets should be long, random, and specific to your installation. Never reuse example values.
Transport encryption
Place a reverse proxy providing TLS in front of the service, or enable HTTPS directly with your certificates. Otherwise, credentials and content are transmitted unencrypted over the network.
Network Exposure
An internal instance does not need to be accessible from the internet. Restrict access to the organization's network, or use controlled remote access.
Accounts and roles
Limit the number of administrators, favor invitations over password creation, suspend rather than delete, require two-factor authentication for administrators, and maintain at least one local administrator account even with single sign-on. A single instance does not isolate multiple companies: roles and storage are global. → Users, roles and organizations
API keys
Each API key grants full programmatic access and applies to the entire instance: neither walls nor per-document permissions apply. Create one per integration, revoke those that are no longer needed, and do not distribute them in shared code. → API developer
What is stored, and where
Everything fits in a single storage directory:
- the application's database — users, spaces, conversations, settings;
- the extracted documents; ;
- vector indices, if the basis is local; ;
- the downloaded models; ;
- the files produced by the agents;
- the audit log, exit logs and security policies.
Protect the storage directory. Encrypt the disk that hosts it and restrict access rights: the database, documents, and indexes reside there. The most sensitive secrets (SSO connectors, email and calendar tokens, two-factor authentication) are also encrypted by the application.
Passwords are never stored in plain text. Provider API keys are hidden from the screen and never displayed again after saving; protect the configuration file as a secret.
Backup and restore
Save this: The storage directory and the configuration file are all that's needed—the application itself will reinstall. Without the configuration file, the encrypted secrets become unreadable.
Recommendations:
- Perform a cold backup, or one that guarantees database consistency.
- Try a restore. A backup that is never restored is not a backup.
- Encrypt your backups: They contain all the documents and conversations.
- Back up your data before each update.
Traceability
The event log stores key actions: logins, account creation and deletion, space modifications, document imports, API key creation, and SSO logins. The audit log keeps a hashed copy of these actions outside the database; deleting the event log does not affect it. → Journals and supervision
An administrator or manager can also view and export the conversation history. Inform users of this: it's a transparency requirement, and often a legal obligation.
No retention period can be set for conversations or the event log: they remain until deleted. The audit log is retained indefinitely, unless you specify a retention period. AUDIT_RETENTION_MONTHS.
Data erasure
| Element | Erasure |
|---|---|
| Conversation | Thread removal. The command /reset It doesn't delete anything: it removes the messages from the history sent to the model, and they remain accessible to administrators and managers. |
| Documents from a space | Removing the space, or resetting the index |
| Full document | Removal from library — removed from all spaces |
| Memoirs | Deleted one by one by the user |
| Second Brain Notes | Individual deletion by the user |
| Account | Deletion also erases conversations and threads; suspension, on the other hand, keeps everything. Guest accounts are deleted upon expiration. |
| Event log | Complete deletion by an administrator |
| Journal of offenses | Complete deletion by an administrator |
| Audit log | No deletion from the application; only AUDIT_RETENTION_MONTHS Removes past due months. It contains account identifiers and metadata |
For a data protection erasure request, all of these elements must be reviewed, including backups and external providers.
Telemetry
Disabled by default: no events are sent until an administrator enables it, in Settings → Privacy and data.
Once activated, the application transmits anonymous usage statistics to guide development. No document content, conversational data, or personally identifiable information is included.
On a sovereignty-required installation, leave it disabled and document it.
Points to be aware of
Browser voice recognition This typically transmits audio to the browser publisher's servers. It's a data output independent of WivenLLM's configuration. It's only used if you select it; the default setting is Whisper locally.
Remote MCP servers They receive what the agent transmits to them. Only connect devices you trust. → MCP Servers
The agent's skills actually make a difference. A file writing skill writes files, a web search skill issues queries. They are activated for the entire instance, not space by space: activate only the minimum necessary.
The web widget is public. The space to which it is linked must contain only distributable information, and the list of authorized domains must be provided: without it, the widget will respond from any site, unless the installation requires this list. → Web widget
The return form, When your version offers it, it transmits your message, screenshots, username, role, and technical information to the publisher. Never your keys, documents, or conversations.
A model can be wrong. Citing sources allows for verification; it does not replace it. For any binding use, human review remains necessary.
Compliance
WivenLLM provides technical tools—data localization, separation of workspaces, roles, logging, and data erasure—several of which need to be activated. Compliance itself (nLPD, GDPR, industry-specific requirements) depends on your configuration and procedures.
The elements to document in your register:
- where the processing takes place, and which subcontractors are involved where applicable;
- which categories of data are indexed; ;
- who has access to what; ;
- the retention periods and the methods of deletion, to be arranged by you for conversations;
- the information given to users.
