AI agent data security, described precisely rather than reassuringly
Most vendor security pages are a list of adjectives. This one is a description of where your data actually sits, what leaves your infrastructure, and what we do about the parts that genuinely have to cross a boundary.
We deploy into your accounts
Your cloud, your databases, your secret store. We do not operate a multi-tenant platform that your data passes through, because that would make us a permanent dependency and an additional place for your records to sit. When an engagement ends, nothing has to be migrated out of anything of ours.
Scoping is enforced at the query layer
Where a system holds data for more than one user, the scope is part of the query rather than a filter applied to results afterwards. One user’s records are never candidates for another user’s search, which is a stronger guarantee than filtering and it is the pattern we used on Synca and on the hosted research platform.
What a model sees is a design decision
Some work genuinely requires sending content to a model provider. We scope that to the minimum the task needs, we tell you exactly what crosses that line before we build it, and where a task can be done by a smaller model running closer to your data, we do that instead. You get the boundary written down, not implied.
- The specific fields sent for each task type
- Which provider and which model receives them
- What is redacted or substituted before sending
- Which tasks run without leaving your infrastructure at all
Least privilege, per agent
Each agent holds credentials scoped to the workflow it runs and nothing else. A support agent that needs to read orders does not hold a key that can also modify pricing. This is unremarkable engineering practice, and it is worth stating because agent projects skip it surprisingly often.
Retention is yours to set
Traces, approvals and retrieval records are useful for debugging and necessary for audit, and they are also data. Retention windows are configuration, they live in your infrastructure, and we will help you set them to match whatever regime you operate under rather than defaulting to keeping everything forever.
Questions about this
Do you train models on our data?
No, and neither do the providers we use under their business terms. Your data is used to answer your questions and for nothing else. Where a provider’s terms matter to your compliance position we will point you at the specific clause rather than paraphrasing it.
Can everything run inside our own network?
What about GDPR and UK data protection?
Who has access during the build?
Send this page to your security lead
If they have questions it does not answer, we would rather have that call early than late.