Skip to content
Crypto.world

The Ledger

Preprint finds wallet checks can miss agents’ spending authority

A new paper says common wallet checks can approve signed objects that permit transfers beyond a buyer’s stated mandate. Its proposed wallet-side checker rejected all 1,680 violating objects in the authors’ test model.

Preprint finds wallet checks can miss agents’ spending authority

A new arXiv preprint says checks commonly used to approve agent payments can miss what a signed wallet object lets its holder do. The authors propose a wallet-side checker that evaluates the full transfer authority carried by an object, rather than relying only on its displayed fields or a simulation of one payment.

Submitted to arXiv on November 18, the paper, “Signed, Valid, and Unauthorized: Authority Closure at the Trust Boundary Between Agent Commerce Protocols and Crypto Wallets,” is by Shengwei You, Aditya Joshi, Andrey Kuehlkamp and Jarek Nabrzyski. It defines “authority closure” as the set of transfer sequences that anyone holding a signed object could enable, and says the object respects a buyer’s mandate only if every possible sequence stays within the buyer’s constraints.

What a signature can permit

The paper’s example involves a buyer authorizing a shopping agent to pay a merchant at most 25 dollars in USDC on Base by Friday. A wallet could display the correct payee, amount, token, network and deadline, while the agent signs an EIP-2612 permit. The permit’s deadline limits when the permit itself can be submitted; it does not necessarily limit when the resulting allowance can be spent.

In the example, the merchant’s key could later be used to move the authorized 25 USDC to any address, the authors say. They argue that checking visible fields or simulating a single payment can miss this continuing authority, as well as transfers enabled through repeated attempts, split amounts or a spender-selected recipient.

The paper says three approaches can fail to establish that every transfer stays within the mandate: matching displayed facts, verifying an object against payment requirements supplied by a merchant, and simulating one execution. In the authors’ analysis, the first approach does not capture how the consuming contract uses the signed object; the second may rely on requirements chosen by a party the buyer does not trust; and the third may not expose authority that persists beyond the simulated transaction.

Checker tested across signed objects

The proposed checker, named Closure, calculates a footprint for each wallet object: its chain and asset, which recipients an executor could choose, the cumulative amount and the time window for transfers. It checks those details against the mandate. The paper says unknown spender code and unknown permission restrictions receive no credit, so the checker fails closed.

The authors evaluated Closure on 2,130 generated objects across 13 families, using 62,979 execution programs. They report that 265 objects with observed violations passed both a field-and-expiry check and an address allow-list. Under the test model, Closure rejected all 1,680 objects with observed violations and accepted the other 450; its median decision time was at most 12 microseconds per family.

The paper also describes two local checkouts using the x402 reference client: approved payments settled, while Closure refused requests outside the mandate before signing. The authors say some authorization forms can meet the mandate when their fields pin the recipient, cap the amount and bound the transfer time; whether a particular object does so depends on its fields and the contracts that consume it.

Limits of the evaluation

They also describe the adversary programs as a finite set, making the observed count of unauthorized objects a lower bound; for objects the checker accepted, the absence of a violation in testing is evidence, not by itself proof.

Sources

  1. Original source · Preprint arXiv arXiv:2611.24989

Related stories

Explore

More articles