Intelligence has a purpose.
Authority has a boundary.
An AI agent needs an owner, a scope and controls it cannot overrule. Explore the separation between reasoning and permission.
Try changing
the request.
Choose a permission on the passport. Then add a persuasive instruction. In this sample, the policy boundary does not change.
This illustrates deterministic permission checks. It is not a claim of perfect injection detection, model alignment or universal enforcement.
Investigate agent misuseSPECIMEN / 086
invoice-assistant
Owner: Finance · Scope: invoice processing
Start with accountable agent identities.
The AI Passport is a proposed policy record, not a visual badge that makes an agent safe. It should identify the owner, intended task, allowed tools and resources, data limits, network destinations and expiry.
Permission changes need an audit trail. Credentials must not be shared across unrelated agents or customers. Higher authority should never follow simply from a model asking for it.
Enforce where the action actually happens.
Requests must pass through a non-bypassable, appropriately authenticated tool boundary. A convenient SDK wrapper is not sufficient if an agent also holds direct production credentials.
Approval should bind to the exact target and operation. “Allow a payment draft” must not become “allow any payment”. Parameters, destination and amount need revalidation immediately before execution.
Treat outside content as data, not authority.
Documents, web pages and tool responses can carry misleading instructions. Content screening may contribute evidence, but it cannot be the sole security boundary.
The planned approach layers least privilege, resource restrictions, input and output validation, data controls, behavioural evidence and human review. These controls still require testing against actual integrations and failure cases.
Model choice should not choose your permissions.
The intended gateway is independent of the model provider. A hosted model, a local model and a customer-selected model should still face the same application-side authority checks.
Model processing, redaction, prompt retention and any external provider must be disclosed before deployment. This website’s demo uses no model and transmits no prompts.
Less noise.
More understanding.
Explore the idea. Challenge the assumptions.
Help shape what comes next.