Ask with a reference
The model sees a capability reference such as cap_weather_read_7f3d. It does not receive the provider credential.
How it works
A request reaches a checkpoint before it reaches a tool.
CanaryNorth gives an AI a narrow way to ask for an action, gives a person a clear decision point, and keeps a record of what happened.
The short versionAsk. Check. Decide. Remember.
The demo is synthetic. The boundary never calls an external tool.
The flow
Four things happen
Check the request
The boundary checks the action, resource, scope, freshness, and content. The request is not trusted just because a model made it.
Pause for a person
Higher-risk changes can wait for a human yes or no. No answer means no action in this synthetic workflow.
Leave a receipt
The record keeps the decision, reason, and references. It does not repeat the secret or the unsafe input.
A useful mental model
The model can knock. It cannot walk through.
Imagine a locked service door. An AI helper can bring a request to the door, but the boundary checks the badge and the request before anything moves. If the request is risky, a person can review it. The receipt is the logbook left at the door.
Try the guided casemodel askspolicy boundarytool only if allowedreceipt
Shown here
What the demo can show
Synthetic capabilities, scoped policy decisions, human approval records, redacted evidence, and tamper-evident receipt shapes.
Still open
What it does not claim
No universal protection, perfect output, live malware scanning, live steganography detection, or real external tool execution. Production identity, key management, monitoring, and independent review remain separate gates.
Next