Glossary

Glossary

Definitions for words used on this site. They record how this reference uses each word. They are not definitions from a standard.

Borrowed Clearance
A control failure in an AI system at a bank: an agent sends money, or performs some other step that cannot be reversed, relying on a check it neither ran nor can confirm. Its rule: Signed, payload-bound clearance is never inherited. Full text on the definition page.
Clearance
As used here, the result of a required check that an action depends on, for example "cleared", "passed", or "approved". A message announcing that result is a report about the clearance, not the clearance.
Inherited control outcome
A check result stated by one part of a system and relied on by a different part that has no means to confirm it. The patterns page describes how this happens.
Payload-bound
Tied to exactly one amount, one payee, and one validity window. A difference in any of the three cancels the approval. See the controls page.
Signed clearance
A record, signed by the service or person who performed the check, that names the payload it applies to. A status sentence quoted in a message does not qualify.
Point of action
The service that carries out the irreversible step, such as executing a transfer. Validation only counts if it happens there, for that specific step.
Fail closed
Treating absent, unconfirmable, expired, or non-matching clearance as a reason to halt and hand the case to a person. Continuing with a warning attached does not meet this standard.
Trust by position
Assuming that output from a peer inside the system has already been checked. The first of the three patterns.
Floating approval
A sign-off filed on a case or supplier record instead of a single transfer, which lets it be reused for another sum or account. Worked through in the reused approval scenario.
Content as authority
Treating a status written inside an attachment, a message, or tool output as a verdict. Worked through in the approval in the inbox scenario.
Audience restriction
In federation guidelines, a rule that an assertion names its intended relying party and that party refuses any assertion aimed elsewhere. NIST SP 800-63C-4 requires it at every Federation Assurance Level. See bearer claims and audience-bound assertions.
Bearer claim
A statement that whoever holds it can present and have accepted, with nothing tying it to the party presenting it. A bare "approved" forwarded between agents behaves this way. RFC 9449 contrasts bearer tokens, usable by any holder of the string, with sender-constrained ones. See bearer claims and audience-bound assertions.
Borrowed authority
An existing phrase about which identity's permissions an action uses. NHI Mgmt Group applies it to an agent exercising permissions that started with a user, a service account, or a connected app. Watchlight AI (18 October 2026) applies it to a request that rides on the permissions of whichever agent relays it. Both differ from Borrowed Clearance; see related terms.
Client-side enforcement
A protection a service depends on but leaves to its caller to perform, so a caller that skips the step gets past it. MITRE lists it as CWE-602 and advises repeating such checks on the server. In this reference's reading, the service that moves the money plays the server's role. See checks that live on the wrong side.
Confused deputy
A program with legitimate authority that is manipulated into exercising it for a party without that authority; the term comes from Norm Hardy's 1988 paper in ACM SIGOPS Operating Systems Review (summary on the related terms page), and Watchlight AI (18 October 2026) applies it to AI agents. In the OWASP Agentic Top 10, entry ASI03 uses the label for agents that accept internal requests without question. See related terms.
Continuous access evaluation
Reassessing access already granted when circumstances change, instead of relying on the original decision until it lapses. CAEP 1.0 defines events such as Session Revoked and Risk Level Change so downstream services can act on the change. See withdrawing clearance when conditions change.
Delegation and impersonation
Two ways one party can act for another, as separated in RFC 8693, OAuth 2.0 Token Exchange. Under impersonation, the actor cannot be told apart from the original subject. Under delegation, it keeps its own identity, named in an act claim, and nested act claims record each earlier hop. See recording who acted on whose behalf.
Object-level authorisation
Checking, for every record a request names, whether this caller may take this action on that record. Its absence heads the OWASP API Security Top 10 2023 as API1:2023, Broken Object Level Authorization. See the OWASP mapping.
Replay
Presenting a genuine message or sign-off a second time to get the same effect again. MITRE covers the network form as CWE-294, with a sequence number or timestamp that can be used only once, and a signature over the message, as mitigations. See replayed sign-offs.
Security decision on untrusted input
A protection that rests on the presence or value of an input that the party being checked is able to set or change. MITRE lists it as CWE-807. See related terms.
Sender-constrained token
A token that the receiving service accepts only from a presenter who proves possession of a key bound to it. RFC 9449 (DPoP, September 2023) does this with a fresh signed proof at each use. See binding a token to its presenter.
Time of check, time of use
The interval in which something that has been checked can change before the result is acted on, so the check no longer describes what is used. MITRE lists it as CWE-367, often shortened to TOCTOU. See time of check, time of use.