- An AI agent can make API calls that are each authorized and still break a business control when those calls are combined.
- Valid tokens and successful API responses don't show that separation of duties or an independent approval actually held.
- Validating an AI agent means following the whole transaction: who started it, which identity ran it, and whether required approvals happened.
Consider a finance team piloting an AI agent that matches supplier invoices to purchase orders and prepares payments for review. It runs under an approved service identity, calls business applications through secured Application Programming Interfaces (APIs), and its permissions have been reviewed.
A routine test then shows the problem. The agent updates a supplier's payment details and then prepares a payment to that same supplier. Both API calls succeed, and neither triggers an access violation. But the finance process says whoever changes a supplier's payment details must not also prepare payments to that supplier. Here, one agent did both.
The example is hypothetical, but the question is real: what happens when every authorization check passes and the transaction still breaks a control?
The Gap Was Between the Calls, Not in Them
Identity controls did their job. The agent was recognized, its token was valid, and each operation was allowed. What no one checked was the relationship between them. NIST describes separation of duty as the principle that no user should hold enough privileges to misuse a system on their own, and notes it can be enforced dynamically, at the moment access happens. That check was missing here.
This isn't really a failure of OAuth or API authentication. Scopes limit what an application can do on a user's behalf, and the resource still has to apply the actual permissions. OAuth scopes alone don't establish whether actions across applications satisfy a business control.
Why Entitlement Reviews Don't Catch It
Access governance asks what a user, role, or service account is entitled to do. The Cloud Security Alliance highlights a related limitation: role-based permissions establish baseline access, but they don't necessarily capture what an autonomous agent does across a sequence of actions.
As Gaurav Sharma stated in his RSAC blog, Why IAM Can't Describe What AI Agents Are Doing, "The same identity may read a finance record during one task and alter it during another." His focus is on task-level authorization. The gap here is narrower and older: a classic separation-of-duties conflict that now sits inside one agent. An entitlement review would confirm our agent can reach both APIs, and the gateway would log two authorized calls. Neither says whether the agent should have combined them on that invoice.
A Human Approval Step Helps Only If It Shows the Right Things
A later approval step doesn't undo that conflict, and it may not even reveal it. If the reviewer only sees an amount and a recommendation, nothing tells them the same agent changed the supplier's bank details earlier. For high-risk transactions, a material change after approval should invalidate it. The decision also has to be enforced outside the language model. A prompt telling the agent not to bypass controls is guidance, not a security boundary.
Testing the Control, Not Just the API
Start with one high-risk workflow, tested from the initial request to the final business action. For an invoice agent, three controls matter most:
1. Transaction-level separation of duties. The policy layer checks whether the same actor or agent has already performed a conflicting action on the same business object.
2. Approval bound to context. Approvers see what changed and which records were affected, and execution is rejected if material details change after approval.
3. End-to-end evidence. Logs connect the initiating user, agent identity, delegated authority, tool calls, business object, policy decisions, and outcome through one transaction ID.
These follow the least-privilege and continuous-verification principles in NIST SP 800-207, applied one layer up, to business workflows.
A Different Acceptance Test for AI Agents
The negative test should run in both directions. First, have the agent change a supplier's payment details and then try to prepare a payment to that supplier. Then reverse it: let the agent prepare the payment and try to change the payment details afterward. Neither order should get past the separation-of-duties control. Each attempt should end in a policy denial or a hand-off to a separately authorized workflow.
Read the results narrowly: they cover one conflict in one workflow. But they tell you whether a critical business rule survived the move from people clicking through applications to agents acting across them. What matters is whether the complete business action was authorized.