Security and sovereignty
Your agents run isolated. You decide what leaves.
Nestor is designed for sensitive files: runs stay on your company's instance, cut off from the network; building only sends the structure of your files; and anything that may be sent is visible and logged.
Isolated runs, on your instance
An approved agent runs on the instance dedicated to your organisation, without calling any model.
- Every run happens in a throwaway container, with no network access and a read-only file system.
- No model on the run path: 0 tokens, and nothing to send.
- In production, the non-isolated run mode is refused.
Building from the structure, not from your data
To write an agent, the model needs to understand the shape of your files, not their content.
- The model receives your request, the names and structure of the files (columns, formats) and made-up rows. Never your real rows.
- Before sending, Nestor searches the package for the real values of your files and blocks the sending if it finds any.
- You see what was sent to the model.
You choose how open it is
Your organisation decides whether an external model and web search are allowed.
- “Internal only” mode: no external model, no web search.
- External model allowed: the text sent goes through automatic masking of some personal data (emails, IBANs, numbers…); what is sent is flagged and logged.
- Web search on: only the search terms leave.
Governance and traceability
It must be possible to review, explain and stop an agent that runs on its own.
- Four eyes: the author of an agent or of a change cannot approve it themselves.
- Version history per agent; a change is replayed on your files and on every past period kept before approval, with a summary for the approver.
- Audit pack per agent and period (versions, approvals, fingerprints of the files read), sealed for integrity.
- After several consecutive failures, automatic runs are suspended.
One instance per company
Nestor is not a shared service: your company has its own instance.
- An instance dedicated to your company, in Europe or on your own infrastructure, with its own data and secrets.
- Inside the company, teams are isolated from one another; each person is an administrator, approver or member.
- Each release ships with its bill of materials (SBOM), a vulnerability scan and a signed manifest, which a script checks before installation.
- Support without access to your data: a diagnostic package (versions, settings without secrets, masked log excerpts) is written on the server; it sends nothing.
- Found a vulnerability? Write to contact@kastorlab.ai (see also /.well-known/security.txt).
What we do not claim
Trust starts with accuracy.
- No certification (SecNumCloud, ISO 27001, HDS…) is claimed to date.
- The audit pack is sealed for integrity; its legal scope is still to be confirmed with a lawyer.
- Some services (Google Drive, Gmail, Outlook, LinkedIn…) are supported but not yet tried on the real service.
- Automatic masking is rule-based: it is pseudonymisation, not anonymisation.
- If you allow an external model, the conversation sends it your messages, the history and the text of attached files.
Frequently asked questions
Security: your questions.
Are my files sent to the AI to build an agent?
No, not their content. The model receives your request, the names and structure of the files (columns, formats) and made-up rows, never your actual rows. Before sending, Nestor searches the package for the real values of your files and blocks the request if it finds any. You see what was sent.
Where is my data hosted?
With the Team and Business offers, on an instance dedicated to your company, in the European Union, managed by KastorLab. With the Sovereign offer, on your own servers or cloud, with your own storage. Nestor is not a service shared between companies.
What is “internal only” mode?
A setting of your organisation: no external model, no web search; only the models installed on your infrastructure answer. This is what we call the closed loop. If you allow an external model, you leave it for the uses concerned, and every outbound request is flagged and logged.
What about the GDPR and the AI Act?
We claim no certification. We prepare compliance with you: documentation, logs, an audit pack per agent and per period, “internal only” mode for the most sensitive data. Automatic masking of personal data is pseudonymisation, not anonymisation.
Who can put an agent into service?
Someone other than its author. The author of an agent or of a change cannot approve it themselves; every version is kept, and a sealed audit pack traces versions, approvals and fingerprints of the files read.
How do I report a security vulnerability?
Write to contact@kastorlab.ai. The site's /.well-known/security.txt file gives the same address.