
Imagine you write a check with a memo line that says 'rent for June, 123 Main St, to Landlord LLC.' Your bank honors the signature and the amount — but it doesn't read the memo. Anyone who gets hold of the check can cash it, because the enforceable part is narrower than the written part. That's the authority gap, and this paper proves it exists in the infrastructure AI agents are starting to use to pay for things. The committed claim: when an AI agent authorizes a stablecoin payment through a commerce protocol like Coinbase's Commerce Payments Protocol, the ERC-20 token contract enforces only a subset of the fields the protocol specifies. The paper defines nine constraint classes — ranging from full binding (the signed data uniquely determines the merchant, order, amount, and facilitator) to essentially no binding (any holder of the signature can redirect funds). They derive these classes for 12 authorization configurations covering the main ERC-20 approval and permit mechanisms. All 31,800 test trials on local copies of deployed contract code confirm the derivations. The empirical punch: of 26,653 sampled USDC authorizations actually redeemed on Base (Coinbase's L2), 98.7% followed a path where any holder of the signature could have redeemed it. The commerce protocol's terms — merchant identity, order ID, facilitator — existed only as unenforced metadata. This isn't a theoretical attack; it's the default state of deployed infrastructure. Architecturally, this is constraint analysis over the ERC-20 token standard's approve/permit/transferFrom call chain. The key insight is structural: ERC-20 approvals bind the spender address and amount, but not the recipient. When a facilitator submits a payment to a plain externally-owned account (no contract code), there's no on-chain entity to check whether the signed terms match the transaction. The gap is a property of the standard, not a bug in any specific implementation. The proposed fix, MandateLock, is a minimal recipient contract that checks signed mandate fields against the actual transaction parameters. It rejected all 10,800 modified mandates in testing and accepted valid signatures from two independent signers, at a cost of 25% more gas than a standard recipient-only transfer. This is a meaningful overhead, but not a dealbreaker — it converts an unenforceable promise into an on-chain constraint. The integrity story is unusually strong for a security paper. The authors test against deployed contract code (not toy implementations), report exact trial counts, sample real on-chain authorizations, and prove sufficient conditions mathematically for the recipient contract to enforce signed mandates. The 98.7% figure comes from actual Base chain data, not simulation. A reproducibility artifact is linked in the manuscript. What makes this matter beyond crypto plumbing: AI agent commerce is the use case. When agents autonomously authorize payments, the gap between what the agent's protocol specifies and what the blockchain enforces becomes a surface for exploitation at machine speed. The paper doesn't just identify the gap — it taxonomizes it, measures it in the wild, and delivers a minimal fix. The question is whether anyone deploys it.
The analogy
You write a check with a detailed memo line — payee, invoice number, purpose. Your bank honors the signature and amount but ignores the memo. Anyone holding the check can cash it. The commerce protocol is the memo; the token contract is the bank. This paper measures the gap between what you wrote and what gets enforced.
Check memo
Detailed instructions written on a check that the bank doesn't read or enforce when processing the payment
Commerce mandate
Merchant, order, and facilitator fields specified by the protocol but not enforced by the ERC-20 token contract
Honest assessment
This is a clean, well-scoped security paper that identifies a real structural problem in how ERC-20 payments work, proves it formally, measures it in the wild, and delivers a minimal fix. The 98.7% figure is the headline: nearly all real USDC authorizations on Base are redeemable by any signature holder. The limitation is that adversarial robustness under real MEV conditions is proven mathematically but not battle-tested. If you're building or evaluating AI agent payment infrastructure, this paper changes what you should require. If you're not, it's a sharp illustration of how 'the protocol says X' and 'the chain enforces X' are very different claims.
What Is CLAIMS?
70
/ 100
BedrockNews applies a structured analytical framework called CLAIMS to academic papers and technical research releases.
How CLAIMS worksClaim
4/5
Ladder
3/5
Architecture
4/5
Integrity
4/5
Milestone
3/5
Successor
3/5
What is this paper actually committing to?
Genuine first: formal taxonomy of the gap between what commerce protocols promise and what ERC-20 signatures enforce. Nine constraint classes, 12 configurations, with a deployed fix. Not incremental.
The field fight
Should AI agent payment security be enforced at the protocol layer or the contract layer?
Side A
Commerce protocols specify enough; wallet UX and off-chain trust handle the rest
Held by: Coinbase Commerce, most deployed dApps
Side B
On-chain enforcement is mandatory; off-chain promises are unauditable at agent speed
Held by: Account abstraction advocates, this paper
Milestone tracker
Unlocks: Autonomous AI agent payments with on-chain guarantee that signed terms are actually enforced
Current
25%
gas overhead
MandateLock immediate payments
Next
≤50%
gas overhead
streaming/recurring mandates
~1-2 years for contract-level extension; multi-year for EIP standard change
Verified Facts (6)
All 31,800 trials on local copies of deployed contract code match the derivation
Of 26,653 sampled USD Coin authorizations redeemed on Base, 98.7% followed a path that any holder of the signature could redeem
MandateLock rejects all 10,800 modified mandates on deployed token code and accepts signatures from two independent signers
MandateLock uses 25% more gas than a transfer that only the recipient can submit
Nine constraint classes defined and derived for 12 authorization configurations
Paper analyzes Coinbase's deployed Commerce Payments Protocol
Practical Implications (4)
Any system deploying AI agents that authorize stablecoin payments should audit whether the signed mandate fields are actually enforced on-chain — this paper shows the default answer is 'they aren't'
MandateLock or equivalent recipient contracts should be evaluated for any commerce flow where the facilitator is a separate entity from the merchant
The 25% gas overhead for MandateLock is a concrete budget number for teams deciding whether to adopt on-chain mandate enforcement
Wallet developers should consider whether displayed recipient information matches what the chain will actually enforce
More latest news
Read all