OWASP

OWASP mapping

Borrowed Clearance is not an OWASP category. It is a name for the point at which published agent-communication and excessive-agency risks meet a payment. The mapping on this page is a reading by this reference.

Rule

Signed, payload-bound clearance is never inherited.

Four OWASP texts are used here: the OWASP Top 10 for Agentic Applications 2026, LLM06:2025 Excessive Agency, the AI Agent Security Cheat Sheet, and API1:2023 Broken Object Level Authorization from the OWASP API Security Top 10. Words in quotation marks are OWASP's own; everything else is a summary and is labelled as such.

ASI07: Insecure Inter-Agent Communication

This entry of the Agentic Top 10 covers messages between agents that are not authenticated, not integrity-protected, or not checked for meaning. Among its examples is replay against trust chains, where resent delegation or trust messages lead agents to grant access or follow outdated instructions (paraphrased).

Its mitigations, summarised, are to sign each message, hash the payload together with its context, look for natural-language instructions that were hidden or altered, and attach nonces, session identifiers, and timestamps limited to the task. The signing control relies on them.

A resent or outdated "approved" is one route by which clearance gets borrowed. ASI07 describes the messaging weakness. It says nothing specific about bank control outcomes and does not contain the rule this site uses.

ASI03: Identity and Privilege Abuse

ASI03 warns that trust between agents, or credentials passed from one to another, can be abused to widen access, take over privileges, or carry out actions nobody authorised (paraphrased).

OWASP presents this entry as the agentic successor to Excessive Agency (LLM06:2025) and notes that it frequently builds on prompt injection (LLM01:2025) (paraphrased). The OWASP page for that second risk is LLM01:2025 Prompt Injection. How the checklist handles a false approval hidden in ingested content is covered on the testing page under C5.

The confused-deputy case described on the related terms page is filed under ASI03. One of OWASP's sample attacks, summarised: a crafted email prompts a mail-sorting agent to tell a finance agent to send money, and the finance agent pays without checking because the instruction came from another internal agent. The example belongs to OWASP. It is neither of this site's worked scenarios nor an account of something that happened.

ASI08: Cascading Failures

ASI08 is the Agentic Top 10 entry about spread rather than origin. In OWASP's description, summarised, one fault (a hallucination, a malicious input, a corrupted tool or poisoned memory) travels from agent to agent and grows into harm across the whole system, because agents plan, keep state and delegate on their own, so a single error can slip past stepwise human checks. OWASP asks that the first defect be filed where it began, for example under ASI07 for a spoofed message, and that ASI08 be used only once that defect fans out across agents, sessions or workflows.

Two of its common examples, paraphrased, are close to this site's subject. In planner and executor coupling, an executor carries out whatever steps a faulty or compromised planner emits, with no validation of its own. In governance drift, human oversight loosens after a run of successes, and bulk approvals or relaxed policies then spread unchecked. One example attack is set in finance: an injected market-analysis agent raises risk limits, and downstream agents trade larger positions while compliance sees only activity "within-parameter".

The listed mitigations include short-lived, task-scoped credentials with every high-impact tool call checked against a policy rule before it runs; an external policy engine that keeps planning apart from execution; checkpoints or human review before high-risk output moves downstream; and tamper-evident, time-stamped logs of inter-agent messages and decisions, with lineage recorded for every propagated action (all paraphrased).

Read alongside Borrowed Clearance, a forwarded "passed" is a fault that cascades well: each agent that accepts it without checking hands it on as if it had been verified, and the payment at the end inherits a check nobody in the chain ran. The point-of-action control is the place where that spread stops. The comparison is this site's reading; ASI08 covers propagation of faults in general, not clearance or payments in particular.

ASI09: Human-Agent Trust Exploitation

ASI09 turns the trust problem around. Here a person trusts an agent, rather than one agent trusting another. OWASP's description, summarised: agents come across as fluent, confident and expert, and attackers or badly designed systems can use that impression to steer what a human decides. The risk grows when people approve actions on the strength of an agent's recommendation or a rationale they cannot verify. OWASP adds that the agent then works as an untraceable "bad influence": the human performs the final, audited step, so the agent's part in the compromise is invisible to forensics. The entry is about human over-reliance; OWASP files an agent's own deviation from its intended purpose under ASI10.

Two of its common examples, paraphrased, are close to this site's subject. When sensitive actions have no final confirmation step, a user's trust turns straight into execution, and one socially engineered prompt can become an irreversible transfer. With fake explainability, the agent invents a persuasive justification that hides its real logic, and the reviewer approves in the belief that the action is sound. Two of the example attacks are set in finance and describe the same case: a poisoned vendor invoice leads a finance copilot to recommend an urgent payment to the attacker's bank details, and the manager approves without independent checks. Another describes consent laundering through "read-only" previews that fire side effects when opened.

The listed mitigations include explicit multi-step confirmation before risky actions; tamper-proof records of user queries and agent actions; a plain-language risk summary that is not generated by the model; verifiable provenance (source identifiers, timestamps and integrity hashes) on recommendations and external data, with runtime checks that block actions lacking trusted provenance or exceeding the agent's declared scope; and comparison of action sequences with approved workflow baselines to catch skipped validation steps (all paraphrased).

Read alongside Borrowed Clearance, ASI09 is the case where the sign-off is real but its basis is borrowed. A person did approve, yet the approval rests on the agent's own account of the facts, so the check the approval seems to stand for was never run independently. This is why the authority control asks whether the signer could clear this payment at all, and why the point-of-action check validates the signed payload rather than the explanation that came with it. The comparison is this site's reading; ASI09 covers human trust in agents in general, not bank clearance.

LLM06:2025 Excessive Agency

Among the excessive-autonomy conditions in LLM06:2025 Excessive Agency is an application that does not verify and approve high-impact actions on its own (paraphrased).

Releasing a transfer because some message said "approved" is exactly such an action, with nobody confirming it at the service that executes. LLM06 covers that gap broadly. Borrowed Clearance narrows it to a bank check the executing agent neither ran nor can confirm.

API1:2023 Broken Object Level Authorization

The first entry of the OWASP API Security Top 10 2023 is older than any agent list, and it describes the same gap one layer down. A caller who is allowed to use an endpoint changes the identifier of the record in the request, and the server acts on a record that caller has no right to touch. OWASP's explanation, summarised: servers often do not keep track of the client's state and instead decide which object to act on from identifiers the client sends.

OWASP's remedy is that every function taking a record identifier from the client must check whether the logged-in user may perform "the requested action on the record". It adds that comparing the session's user ID with the identifier in the request covers only a small subset of cases. Its examples include a vehicle API that accepted any vehicle identification number without confirming that the vehicle belonged to the caller, and a delete call that removed documents by ID "without any further permission checks". It also asks for tests of the authorization mechanism, and for changes that fail them not to be deployed.

Read alongside Borrowed Clearance, an approval that is stored or forwarded with only a loose reference works like an object ID chosen by the caller. The executing agent sees a valid session and a real approval record, but nothing ties that record to this payee, this amount, and this moment. The question API1 asks of every record, whether this caller may take this action on this object, is the question the binding control asks of every approval. The parallel is this site's reading. API1 is about API authorization in general and does not mention agents or payments.

Side by side

OWASP entry Scope Role on this site
ASI07: Insecure Inter-Agent Communication Messages between agents that lack authentication, signing, or replay protection Basis for signing payload and context and limiting exchanges to the task window
ASI03: Identity and Privilege Abuse Trust and credential hand-down between agents, confused-deputy cases included The adjacent authority problem, kept separate from the clearance problem
ASI08: Cascading Failures One fault spreading across agents, sessions and workflows Why an unchecked "passed" grows as it is forwarded, and why the check sits at the point of action
ASI09: Human-Agent Trust Exploitation People approving on an agent's confident recommendation or unverifiable rationale Why a human sign-off built on the agent's own account is not an independent check
LLM06:2025 Excessive Agency High-impact actions taken without independent verification and approval Basis for validating where the money moves, and for halting by default
API1:2023 Broken Object Level Authorization Records acted on by client-supplied ID without a per-object authorization check Older, non-agent precedent for checking each approval against the exact action it covers
AI Agent Security Cheat Sheet Approvals bound to one action, replay protection, and what a signature does not prove Background for the five controls

The cheat-sheet points are covered on the controls page and are not restated here. OWASP has no entry called Borrowed Clearance, and the table above is this site's reading, not an OWASP classification.