Section 8.2 of a new authorization proposal contains a sentence that, on first reading, looks like something left out:
"This specification does not depend on an authenticated identifier that is unique to AI Agents."
The proposal is AAuth, an individual Internet-Draft that the Internet Engineering Task Force (IETF) has not adopted. Its second revision appeared November 1. A document about governing AI agents that says it has no need to know whether the caller is one seems to have misplaced its subject.
The sentence makes sense once you notice that "who is this agent?" packs three questions together, and a different control answers each:
- Which software sent this request?
- What may it do without asking anyone?
- Did a person approve this particular action?
The first is the cheapest to answer, so systems usually start there. Then they let that answer stand in for the other two.
A separate IETF working-group draft on signed automated traffic takes up the first question. The software publishes a key at a web address and signs its requests with it. A server that verifies the signature learns that the request was signed by whoever holds that key, which gives it an identifier it can log, rate-limit, allow or block. Who operates the software, and whether its traffic is benign, the server still has to judge for itself. The draft says plainly that a valid signature proves nothing about whether a request is permitted.
AAuth works on the second question. It builds on OAuth, the standard that lets an app reach your account without being handed your password. OAuth calls a named bundle of permissions a "scope." AAuth defines a deliberately small one, ai-agent-minimal, covering only operations the service judges "completely safe." Whoever sets up the agent can grant it that scope and nothing more. Requests outside it are refused, and the refusal comes back carrying a request for more access.
That is why Section 8.2 can set agent identity aside. The limit comes from what was granted, so learning that the caller is an agent would add nothing to it. The draft still requires a registered, authenticated client and turns away public clients. What it declines to ask is whether that client is an AI.
The weak point is the phrase "completely safe," which records the service's own judgment about harm. Like any default, it deserves scrutiny, since the people who will rely on it are not the ones who set it.
On the third question, the draft is candid. After a person approves, the agent receives temporary permissions that can cover several calls, and the authors list binding an approval to one exact operation as an open problem. In practice, you ask an assistant to pay one invoice, you tap approve, and the software may now hold permission to make payments in general. An existing OAuth extension, RFC 9396, lets a request spell out the amount, currency and recipient, so that what the person approves is a transaction rather than a capability. Even then, each API still has to check that the payment it executes matches the one that was approved.
Where nothing makes that check, the person tapping the button is the only link between what they meant and what the agent did. A 2025 government advisory on the criminal group Scattered Spider describes attackers sending sign-in prompts repeatedly until employees pressed "Accept." Those were login prompts, not agent approvals, but the habit the attackers exploited, approving whatever makes the prompt go away, is one agent approvals will meet too. Organizations also have their own reason to keep the click: it produces a record that someone said yes, whether or not the yes constrained anything.
When you read an agent design, ask which of the three questions each control actually answers. Then test the third one directly. Approve a payment, then change the recipient, raise the amount, and try a second payment under the same grant. If any of those goes through, the approval recorded that you were present without limiting what followed.
-
Matching approval to request: OpenID's CIBA specification defines an optional
binding_message, a short code shown on both the device that started the request and the device that approves it, so a person can connect the two (though nothing guarantees they will look). -
When consent actually happens: In the authorization-code flow of RFC 9396, a person consents during the authorization request, and the later token request is checked against that grant instead of asking anyone again.
-
Prompt limits versus enforcement: The EMNLP industry paper "Prompts Don't Protect" found unauthorized tool calls ranging from 4% to 37% when allowlists lived only in the prompt, and none when a proxy rechecked authorization at each invocation, though only in a controlled test environment.
-
Scopes shaping tool visibility: The current MCP tools specification lets a server return only the tools a caller's scopes permit, which moves part of the standing-access question into discovery itself.

