Patterns
How it happens
Borrowed Clearance becomes likely when a control outcome is trusted because of where it came from, because it was stored against the wrong object, or because a status line inside untrusted content is read as a decision.
Rule
Signed, payload-bound clearance is never inherited.
The account published with the term pictures a bank that splits payment work across several agents. One gathers the request, another runs or collects checks, and a third executes the transfer, with tools in between and messages passing along the chain. Any of those messages can contain a status such as a sanctions screen "cleared" or an ops sign-off "approved". For the agent that executes, such a status is only words on a page. Yet that agent controls the transfer tool.
Three recurring shapes turn those words into something the executing agent treats as clearance.
Trust by position
The executing agent assumes that anything sent by a peer inside the system has been checked already. Where the sender sits in the architecture stands in for evidence the executing agent could inspect for itself. Sitting inside the perimeter proves nothing about whether a know-your-customer review, a sanctions screen, a fraud score, or a sign-off ever happened for this particular transfer.
Approvals that float
A sign-off is filed on a case or a supplier record instead of on a single transfer. Since nothing ties it to one amount, one payee, and one validity window, it can be picked up again. A later transfer, even one for a different sum or to a different account, can lean on that earlier record. The reused approval scenario works through this shape.
Content as authority
A status written inside something the system merely reads, such as an attachment, a customer message, or the output of a tool, gets treated as a verdict. It might claim that a review succeeded or that a release is authorised. Whoever reads it never ran that review, and handing the sentence onward does not turn it into clearance. The approval in the inbox scenario works through this shape.
Bearer claims and audience-bound assertions
The three patterns above succeed when a status line can be presented anywhere, by anyone who holds it, with nothing that names the intended consumer. Digital identity federation already draws that line for authentication assertions, and the comparison is useful even though the guideline is not about bank payments or AI agents.
NIST Special Publication 800-63C-4 (Digital Identity Guidelines: Federation and Assertions, published 1 August 2025) requires audience restriction at every Federation Assurance Level: the assertion must target a specific relying party, and that party must confirm it is the intended audience before accepting it (NIST SP 800-63C-4, paraphrased; DOI 10.6028/NIST.SP.800-63c-4). The same guideline distinguishes presentation modes. A bearer assertion can be presented on its own as proof; capturing a valid one can let an attacker impersonate the subscriber at the relying party. A holder-of-key assertion instead names an authenticator the subscriber must prove possession of to the relying party, so stealing the assertion alone is not enough (same source, paraphrased).
Read against this site's patterns, a floating "approved" or a status copied from content behaves like a bearer claim with no audience: any later agent can present it, and nothing in the claim says which transfer service or which payload it was meant for. An audience-bound, payload-bound clearance that the payment service verifies for itself is the opposite shape. NIST SP 800-63C-4 does not discuss AI agents or payment approvals; drawing the parallel is this reference's reading.
Time of check, time of use
Software security has a long-standing name for a check whose result is trusted after the thing it examined may have changed. MITRE's Common Weakness Enumeration lists it as CWE-367, the time-of-check time-of-use (TOCTOU) race condition: a product checks the state of a resource before using it, but that state can change between the check and the use in a way that invalidates the result of the check (CWE-367, paraphrased). The entry's classic example is a privileged program that confirms a user may write to a file name, then opens that name, by which point an attacker may have swapped what the name refers to. Among the mitigations it lists are taking a lock before the check rather than after it, so that the resource as checked is the same one as used, and rechecking the resource after the use call (same source, paraphrased).
The agent version of the gap is wider, though the shape is the same. A sanctions screen or a sign-off is a check made at one moment against one set of facts. The transfer happens later, possibly after the payee, the amount or the account behind a name has changed, and possibly in a different agent that never saw what was examined. A stored or forwarded "cleared" is a time-of-check result carried to a time of use that nobody rechecked. Binding the approval to the exact payload and verifying it where the money moves, as the controls describe, plays the role that locking before the check plays in the CWE entry. CWE-367 is written about files, memory and threads, not AI agents or payments; the comparison is this reference's reading.
Replayed sign-offs
A genuine sign-off can also be sent twice. MITRE's Common Weakness Enumeration files the network form of this under CWE-294, authentication bypass by capture-replay: the design lets someone who has observed a valid message resend it, unchanged or with small edits, and get the same effect as the original (CWE-294, paraphrased). The entry notes that such attacks are common and hard to stop without cryptography. Its two mitigations work together. The first adds a sequence number or timestamp, covered by a checksum, so that each message can be accepted only once. The second signs messages, since anyone who can watch the traffic can also see the sequence numbers and would otherwise just edit them along with the content (same source, paraphrased).
In an agent pipeline the observer need not be an attacker on the wire. Any component that has seen an earlier "cleared" can pass it on again: a retry loop, a memory store, a summary that gets re-read, or a later task that inherits an earlier one's context. Every copy is as genuine as the first, so a check that only asks whether the sign-off is real will accept every one of them. What stops the second transfer is a property of its own: a single-use identifier the executing service records once it has been spent, and a validity window, both inside the signed payload. That is the replay half of the signing and verification controls, and the OWASP mapping summarises the agentic guidance on nonces and timestamps. CWE-294 deals with authentication traffic, not with AI agents or payments; the comparison is this reference's reading.
The executing agent's blind spot
In each shape the agent that moves the money performed no check of its own. The evidence is out of its sight, and it has no way to separate a genuine sign-off from one that was invented, expired, resent, or made up by a model. The controls are designed so that a transfer never rests on that blind spot. The OWASP mapping lists the published risk entries closest to this failure, which cover unsafe agent messaging and trust between agents. They border the term rather than rename it.