Controls
Controls and mitigations
The rule belongs in the architecture. A prompt that tells an agent not to inherit clearance is not a control. The five controls below are how the rule is applied to a payment.
Rule
Signed, payload-bound clearance is never inherited.
For a payment, the rule turns into five checks that the system enforces. They sit beside authority controls, not in place of them. An authority control settles which identity's permissions an action uses. A clearance control settles which check the action is allowed to rely on. Passing the first says nothing about the second: an agent with every permission it needs can still release funds on a "passed" it never confirmed. The related terms page draws that line in more detail.
1. Tie each approval to a single payload
An approval names one amount, one payee, and one validity window. If the amount, the payee, or the window differs from what was approved, the approval no longer applies. Storing "approved" on a case file or a supplier record breaks this control, and that is exactly how the floating approval pattern starts.
OWASP's AI Agent Security Cheat Sheet makes the same demand: approval should be bound to the exact action, and the approval record should hold the actor, the tool, the target, the normalised parameters, a timestamp, and an expiry (OWASP Cheat Sheet Series, AI Agent Security, paraphrased).
2. Have the checker sign it
Clearance travels as a signed record produced by whoever performed the check, whether that is a screening service or a named reviewer. A chat line or a forwarded note reporting that a review succeeded is a claim about clearance, and passing it along does not add a signature to it.
For messages between agents, the OWASP Top 10 for Agentic Applications 2026 (entry ASI07) recommends signing messages and hashing the payload together with its context, checking for altered or hidden natural-language instructions, and protecting every exchange with nonces, session identifiers, and timestamps scoped to the task (OWASP Top 10 for Agentic Applications 2026, paraphrased).
3. Check it where the money moves
The service that executes the transfer validates the signature and the payload binding on its own, for each transfer. A check performed upstream by some other service does not count for this step. The cheat sheet recommends short-lived authorisation artefacts with replay protection for operations that cannot be reversed (OWASP Cheat Sheet Series, AI Agent Security, paraphrased).
It also keeps proposing and executing apart. An agent may suggest a transfer, but a policy service or the execution component should confirm scope, privilege, and approval state independently before anything runs (same source, paraphrased).
4. Stop when in doubt
If clearance is absent, cannot be validated, has expired, or describes a different payload, the transfer halts and a person takes the case. Carrying on while raising a flag is not acceptable, because a flag is not the check that was skipped.
The cheat sheet applies the same default to the execution path: if risk classification, approval validation, policy lookup, or audit logging fails, the action should not go ahead (OWASP Cheat Sheet Series, AI Agent Security, paraphrased).
5. Confirm the signer was entitled to clear it
A valid signature proves origin. It does not prove that the origin had the right to approve this transfer. The cheat sheet says so directly: a correct message signature is not, on its own, permission for the requested action (OWASP Cheat Sheet Series, AI Agent Security, paraphrased). So the verifier also confirms that the signer holds the right to approve the amount and payee in the payload. Here the clearance check and the authority check meet, and both must pass. More on that boundary under related terms.
Precedents in payment and authorisation standards
Binding approval to one payload is not a new idea invented for agents. Two published standards already apply it to payments made by people and by client software, and they show what the first and third controls look like when written as rules.
Under the European rules for strong customer authentication, the requirement called dynamic linking says the payer must be shown the amount and the payee, and the authentication code produced must be specific to that amount and that payee. The provider may only accept a code that matches the original amount and payee, and any change to either one invalidates the code (Commission Delegated Regulation (EU) 2018/389, Article 5, paraphrased; the linked text is the version retained in UK law). In this reference's terms, the code is a clearance that cannot float to a different transfer.
OAuth 2.0 Rich Authorization Requests (RFC 9396, May 2023) make a similar move for API access. The RFC observes that a coarse scope cannot express a request such as transferring 45 euros to one named merchant, and adds an authorization_details parameter whose payment example carries the instructed amount, the creditor's name and the creditor's account (RFC 9396, paraphrased).
Neither standard mentions AI agents, and linking them to Borrowed Clearance is this reference's reading. What they add is evidence that payload-bound approval is already an established, implementable pattern in payments. An agent pipeline that passes around a bare "approved" is falling short of practice that exists outside it.
Recording who acted on whose behalf
The fifth control asks whether the signer was entitled to clear the transfer. Answering that needs a record of who is acting and for whom, and OAuth 2.0 Token Exchange (RFC 8693, January 2020) already defines one for tokens passed between services.
The RFC separates two arrangements. Under impersonation, a party holding the token is treated as the original subject, so anyone receiving it believes it is dealing with that subject. Under delegation, the acting party keeps its own identity and is understood to be acting for the subject; the RFC remarks that the actor is, in a sense, an agent for the subject. Delegation is written into a token with an act claim naming the current actor, and a chain of hand-offs is expressed by nesting act claims, the most deeply nested being the earliest actor. A separate may_act claim states which party is allowed to become the actor for a given subject (RFC 8693, sections 1.1, 4.1 and 4.4, paraphrased).
Read against this site's patterns, an agent that forwards a bare "approved" works like impersonation: the executing agent sees only the conclusion and cannot tell which step produced it. A nested actor chain is the opposite, because every hop stays visible to the party that moves the money. The RFC also limits what the act claim does: it identifies actors and does not by itself make a token valid, so it supports the fifth control rather than replacing the signature and payload checks above. RFC 8693 does not mention AI agents or payments approvals; the comparison is this reference's reading.
Binding a token to its presenter
The second and third controls ask that clearance travel as something the executing service can check for itself, not as a line anyone can forward. OAuth 2.0 Demonstrating Proof of Possession (RFC 9449, September 2023) applies a closely related shape to access tokens.
DPoP is an application-level mechanism that sender-constrains OAuth access and refresh tokens. The client proves it holds a public/private key pair by sending a DPoP proof JWT in an HTTP header; the authorisation server binds the issued token to that public key; and the resource server accepts the token only when a fresh proof shows the same key is still in the client's hands (RFC 9449, sections 1 and 2, paraphrased). Sender-constrained tokens stand in contrast to a typical bearer token, which any party that holds the string can use (same source, paraphrased). The RFC's primary aim is to stop leaked or stolen tokens from being usable by anyone else, by requiring possession of the private key at every use (paraphrased).
Read against this site's patterns, a bare "approved" passed between agents behaves like a bearer claim: whoever holds the sentence can present it. A signed clearance that the payment service verifies against a known checker, for a named payload, is the opposite shape, closer to a proof that only the entitled party can make. DPoP itself is about authority tokens, not bank checks, and the RFC does not mention AI agents or payment approvals; drawing the parallel is this reference's reading. What it adds to the controls above is a published, implementable pattern for refusing a forwarded artefact that the presenter cannot prove they are entitled to use.
Withdrawing clearance when conditions change
A sign-off that was valid when it was given can stop being valid before the money moves: the reviewer's role changes, a device falls out of compliance, or a risk score jumps. The first control pins clearance to one payload; this section is about pinning it to a moment as well. The OpenID Foundation's Continuous Access Evaluation Profile (CAEP 1.0, a final specification built on the OpenID Shared Signals Framework) offers a published model for telling downstream services that an earlier decision no longer holds.
CAEP defines event types that a transmitter sends to cooperating receivers so they can reduce access already granted to human or robotic users, devices, sessions and applications. Among them are Session Revoked, Token Claims Change (a claim inside an issued token now has a new value), Credential Change, Assurance Level Change, Device Compliance Change and Risk Level Change. Each event can say what triggered it, whether an administrator, the user, a policy evaluation or the platform itself (CAEP 1.0, sections 1 to 3, paraphrased).
Read against this site's patterns, a stored "approved" with no route for withdrawal is a decision frozen at the moment it was made. An agent that releases a payment days later is relying on a check whose conditions may have lapsed, which is the reused approval scenario seen from the other side. A receiver that listens for change events, and treats a relevant one as cancelling clearance it holds, fails closed in the sense of the fourth control. CAEP is written for sessions and access tokens; it does not mention AI agents or payment approvals, and the comparison is this reference's reading.
Checks that live on the wrong side
The third control asks the service that moves the money to check clearance itself. A long-standing software weakness entry describes what goes wrong when it does not. MITRE's Common Weakness Enumeration lists it as CWE-602, Client-Side Enforcement of Server-Side Security: a server that relies on its client to carry out a protection meant to protect the server (CWE-602, Description, paraphrased).
The entry's first example is small and familiar. A client asks the server to authenticate a user and only sends an address change if the reply says the login worked. The server answers the login request, then accepts the change command on its own, without checking whether that same session ever authenticated. Anyone who writes their own client can skip the first step (same source, Demonstrative Example 1, paraphrased).
The listed mitigations follow the same line as this page. Any security check run on the client should be repeated on the server, since an attacker can change values after the check or remove the check altogether. Where the two sides must trust each other to some degree, the entry advises integrity checking and strong authentication of inputs, with that trust managed in one central place so that a single code path does not quietly miss a check (same source, Potential Mitigations, paraphrased).
Read against this site's patterns, an executing agent that releases a payment because an upstream agent reported "approved" is the server in that example: the check happened somewhere it does not control, and nothing at the point of action repeats it. CWE-602 is written about client and server software, not AI agents or payments; the comparison is this reference's reading.
A signature is only as good as its verifier
The second control asks for a signed clearance record, and the third asks the payment service to check it. Signed records are often carried as JSON Web Tokens, and the IETF's best-practice document for them shows how a check can look present while doing nothing. JSON Web Token Best Current Practices (RFC 8725, BCP 225, February 2020) lists the failures and the counter-measures (RFC 8725, paraphrased throughout).
Several failures sit in the verifier rather than the signer. Some libraries believed a token's own header when it named the algorithm "none", and so "validated" it without looking at any signature; others could be tricked into checking an RSA-signed token as if it were HMAC, using the public key as the shared secret (section 2.1). The fixes are to let the caller fix the set of acceptable algorithms, use each key with exactly one algorithm, and reject the whole token if any cryptographic step fails, including the inner signature of a nested token (sections 3.1 and 3.3).
Other failures come from accepting a genuine token in the wrong place. A substitution attack takes a token legitimately issued for one recipient and presents it to another; cross-JWT confusion is the variant where a token minted for one purpose is used for a different one (sections 2.7 and 2.8). The document asks the recipient to confirm that the signing keys belong to the stated issuer, to reject any token whose audience is missing or names someone else, to give each kind of token an explicit type, and to write validation rules for different kinds so that one kind can never pass as another (sections 3.8, 3.9, 3.11 and 3.12).
Read against this site's controls, a payment service that accepts a reviewer's signed sign-off without pinning the algorithm, the issuer's key, the audience and the token type has a signature on file but no verification. A sign-off issued for a screening workflow, a different service or a different kind of decision should fail at the point of action rather than pass because its signature is valid. RFC 8725 is about tokens in general, not AI agents or payments; the comparison is this reference's reading.
A one-line test, and when to repeat it
The working test for the term: a payments agent that cannot prove its clearance should be treated as having none.
Neither family of control is optional. Drop the authority side or the clearance side and a transfer can still find a route through. Teams should red-team the full chain before launch and repeat the exercise whenever a tool, an agent, or a hand-off between them is changed. The testing page suggests where those cases fit in a published banking checklist, whose re-test triggers include changes to tools or agent permissions.