OWOnly What It NeedsAgent Security Lab
Synthetic experiment · Policy version invoice-overdue-v1

Give the agent the task—not the whole file.

Watch one trained agent request only the facts it needs. An independent policy gateway—not the model—decides what crosses the boundary.

Run the invoice experiment ↓
ACCESS MANIFEST · YAML
task: invoice-overdue
allow:
  - invoice_date
  - due_date
  - payment_status
deny:
  - bank_details
  - customer_address
reads: 1
memory: discard
The agent can ask.
The policy decides.
ONE TASK→EXACT FIELDS→ONE READ→VISIBLE RECEIPT
Public test case 01

Is this fictional invoice overdue?

The complete record has seven fields. The useful answer requires three.

01 · AGENT REQUEST

Waiting to begin

Choose a scenario and run the policy.

No fields requested yet
02 · POLICY GATEWAY

Independent decision

The model cannot change this rule.

Policy invoice-overdue-v1Maximum reads 1
03 · TASK RESULT

Not started

Nothing has crossed the boundary.

0 fields released7 withheld
Ready for a deterministic replay.
What this proves

The policy can refuse the agent.