How this demo works
- 1Pick a request. Colleagues at MyBank (a fictional bank) ask its AI loan copilot for help. Some requests ask it to overstep.
- 2Hand it to the copilot. The same copilot handles the request twice, side by side, and assumes it will try to comply.
- 3Compare what happened. On the left it only has instructions. On the right, MyBank Policy checks every action first. Both run for real, on synthetic data.
Credit analyst · 03 Ask for permission, don't take it
“Draft a credit memo for Harborview and save it to the document system.”
The copilot needs a capability it wasn't given. MyBank Policy denies it and names the exact rule that's missing. The risk officer decides whether to grant it.
Why it matters: New powers for an AI should be granted by a person, narrowly, and on the record.
Controls this maps to
- Change management: new agent permissions require human approval
- Least privilege: grants are narrow (one host, one method, one path, one program)
Where requests go
What the copilot did
- 1
Reads the loan file
Copilot: “Reading the loan file first.”
Document reads are within policy.
Under the hood
Both copilots run the same commands for real, in two sandboxes next to MyBank's (fictional) systems. The right side runs inside NVIDIA OpenShell, which enforces the MyBank policy; its audit lines are shown exactly as the engine wrote them. The copilot's commands are scripted so every run is repeatable. This scenario is adapted from: