x402 is a standard for paying for a web request with a crypto payment. A seller answers a request with a payment request (HTTP status 402), the buyer pays, and the seller delivers. Because money moves on one side and a web server answers on the other, security researchers have started to look at where the two meet. This page summarizes what three public preprints found, what a buyer can check before paying, and which of those risks our checks do not cover. We are not security researchers and this is not a security audit of any seller.
What the papers found
All three are arXiv preprints. None is described on its arXiv page as peer reviewed, and we have not reproduced their results.
Five attacks on the x402 protocol. In Five Attacks on x402 Agentic Payment Protocol (Zelin Li, Qin Wang, Zhipeng Wang, May 2026), the authors argue that joining a fast web request to a slower blockchain settlement creates gaps. They describe five attacks, say all five are practical, and report auditing three open-source software kits and their endpoints:
- Revert-grant. The seller delivers before the payment is final, so a chain reorganization or replacement can undo the payment afterward.
- Settlement preemption. Someone who sees the payment header settles it first, so the buyer pays but the seller's own settlement fails and the buyer is denied.
- Replay. One payment is reused to unlock several responses, because the seller does not record it as used before delivering.
- Proxy and cache confusion. Proxies or caches alter the payment header or store paid responses, so people who did not pay can receive paid content.
- Server selection. Attackers manipulate discovery listings or flood them with fake servers, steering agents toward malicious paid endpoints.
Four flaw classes. In Free-Riding the Agentic Web: A Systematic Security Analysis of x402 Payments (Shengchen Ling, Yihang Huang, Yuefeng Du, Yuan Chen, Yajin Zhou, Lei Wu, Cong Wang, May 2026, revised June 2026), the authors name four classes: a payment for one resource unlocking another of equal price, concurrent requests all being served on a single payment, spending beyond an approved allowance, and a settlement rate limit that leaves requests delivered but never paid. They report that these let a requester receive resources without paying, up to 100 percent in their tests against official software kits and a production deployment, and say they disclosed the findings.
Metadata privacy. Hardening x402: PII-Safe Agentic Payments via Pre-Execution Metadata Filtering (Vladimir Stantchev, April 2026, revised June 2026) looks at a different risk. Payment requests carry resource URLs, descriptions and reason strings to the seller and to a payment service before settlement, and those can contain personal data. The author proposes a filter that removes it and reports an F1 score of about 0.9 on a synthetic test set.
What a buyer can check before paying
- Make an unpaid request and read the 402 response. A readable price, network, token and recipient address is the minimum.
- Confirm the recipient address matches what the seller publishes, and see whether many unrelated listings share it.
- Prefer sellers that settle before they deliver, and that use a payment service you recognize.
- Start with a small payment and check that you received what was promised.
- Keep payment descriptions free of personal data.
What our checks help with
- Recipient attribution and the shared-address rule. We tie a payment address to the listings that declare it. When many listings share one address, no single agent is credited. This helps against fake or duplicated listings, which is the territory of the server-selection attack, but it does not detect a malicious seller with a unique address.
- Readable 402 validation. Our free checks send unpaid requests only and show whether a buyer would see a valid payment request.
- Settlement verification. For payments on Base we look for the on-chain record that a payment authorization was used and USDC moved, so a claimed payment has to show up on the chain.
- Live checks. We re-test whether an endpoint answers, and keep dated history.
- Upgrade readiness. We note whether a seller is prepared for coming network changes. This is about staying compatible, not about the attacks above.
What our checks do not cover
- We do not test for replay, concurrent-request races, allowance overdraft, cache behavior or header handling. A seller that passes our checks may still have these flaws.
- Our small daily paid tests make one purchase. They show a seller delivered once; they cannot show it would refuse a reused payment.
- We do not audit the payment services (facilitators) or the software kits sellers use.
- We do not look at what personal data your payment descriptions expose.
- Reorganization risk, the revert-grant case, depends on how many confirmations a seller waits for. We do not measure that.
- A passing check is a dated snapshot, not a statement that a seller is safe.
Sources
- Z. Li, Q. Wang, Z. Wang, Five Attacks on x402 Agentic Payment Protocol, arXiv:2605.11781, 2026.
- S. Ling, Y. Huang, Y. Du, Y. Chen, Y. Zhou, L. Wu, C. Wang, Free-Riding the Agentic Web: A Systematic Security Analysis of x402 Payments, arXiv:2605.30998, 2026.
- V. Stantchev, Hardening x402: PII-Safe Agentic Payments via Pre-Execution Metadata Filtering, arXiv:2604.11430, 2026.
Related: test your endpoint for free and how we score readiness.