Most teams scope an AI agent by its job title:
“Invoice processing agent.”
“Customer support agent.”
“Research agent.”
That’s too broad.
The right question is: what is the most damaging action this agent could take, and what must be true before it can take it?
Consider a hypothetical invoice agent. Its task is to read invoices, match them to purchase orders, and prepare payments.
It has access to the procurement system, vendor records, and payment workflow. Then an invoice contains hidden instructions telling the agent to update the vendor’s bank details before processing the payment.
The agent isn’t “hacked” in the cinematic sense. It is following text that entered its working context. The failure is more basic: a low-trust input was allowed to influence a high-consequence action.
The agent’s task was legitimate. Its authority was not properly bounded.
Scope authority by consequence
Reading an invoice and changing a bank account are different classes of action.…