Uncontrolled disclosure
Sensitive content reaching a model or place it was never approved for.
Architecture
Built to be inspected.
Plain language first, technical detail second. This is the path a request takes through Vrangr and the control enforced at each step. A trustworthy control layer should be easy to check.
Data flow
One path, five controls.
Source
Protected data
Sensitive records are tagged when they come in, so the system knows what each one holds.
Check
Identity and access
Each request is tied to a known person and resolved against role and rules before anything is fetched.
Control
Controlled retrieval and redaction
Only approved context is retrieved, and restricted detail is removed or summarized before it reaches a model.
Route
Approved routes
The request runs only in environments and models cleared for that class of data.
Record
Output controls and audit
The answer is checked against export rules before it leaves, and recorded with a clear trail.
What it is designed against
The risks Vrangr controls.
R-01
R-02
Over-retrieval
A search pulling beyond what the requester is allowed to see.
R-03
Boundary failure
Restricted material crossing into a place that should not hold it.
R-04
No record
An answer that cannot be reviewed or explained after the fact.
The controls in detail
Rules
Retrieval and routing decisions are made against an explicit set of rules. The rules are resolved before anything is fetched, so the system never pulls content the requester is not allowed to see.
Identity and access
Every request carries a known identity, resolved to a role and a set of permissions. Access to a class of data depends on who is asking, not on where they sit on the network.
Routing
Each class of data is tied to a set of approved environments and models. A request is never sent to a place outside that set.
Output controls
Generated answers are checked against export rules before they leave the boundary, and recorded with enough detail to review them later.
Audit trails
What was asked, what the rules decided, what was retrieved, where it ran, and what came back are kept as a clear chain for each request.
Limits and non-goals
What Vrangr does not claim.
Stated plainly, because a control layer that overstates itself is not trustworthy.
- ·
It does not remove all risk
Vrangr reduces and controls exposure. It does not make protected data risk-free.
- ·
It does not replace your governance
Vrangr enforces rules. Deciding what those rules should be stays with your team.
- ·
It does not vouch for the model
Vrangr controls what a model sees and where it runs. It does not guarantee the answer is correct.