Security reviews we did in the open, on other people’s specs, code and tools: on GitHub, on forums and on Reddit. Each one is listed because its author confirmed the point or changed something, and each links to the place you can check it.
49 security reviews the author confirmed: 7 on GitHub, 27 on forums, 15 on Reddit.
These are not audits, and nobody paid for them. Full audits are in audit reports and case studies.
Showed that the delegated-authority registration records a per-transaction cap without its unit, so one sample record reads as 10 refused under a cap of 5000.
Pointed out that an x402 upto settlement of 0 sends no transaction, so a reader could free a hold whose signed payment can still be collected until its deadline.
Showed that a standing "compensate: allow" grant had no amount limit and that money sent out is not reversible by the sender; the draft now has spending totals per run and per period.
Reported that a human's approval covered the tool arguments but not the payment terms the server sets; the next release binds those terms into the approved action.
Pointed out that readers had no rule for a token storing both metadata keys or both decimal fields, and that metadata can be rewritten after listing; the author added a precedence section and a security concern to the draft.
Wrote a worked case of one x402 payment through MCP's messages, showing that a payment re-signed to a different payee passes as a valid retry; it became test vectors CS41 to CS49.
Checked the three new IBC light-client queries against ibc-go v10.7.0 and pointed out that contracts verifying proofs against ConsensusState must check client status and pin proof specs themselves; the author added that to the changelog.
Pointed out that a magic-value opt-in answers for the contract code rather than the account, so one EIP-7702 delegate or clone implementation opts in every account behind it; the author confirmed it.
Pointed out that Lindell17 two-party signing lets the other device extract the finishing device's key share through crafted failed signatures (CVE-2023-33242); the author accepted it and will move the design to DKLs23.
Showed that the reference policy wrote new storage slots inside a fixed gas budget, that the example escrow let the payer pick the payee, and that claims trusted the policy; the author fixed all three.
Showed that malformed return data from one leaf reverted the whole composite evaluation, so a single broken verifier refused subjects on the allowlist; the author reproduced it and fixed it in v0.1.2.
Pointed out that the check missed tokens on AccessControl, where a role admin controls upgrades, and Safe modules that bypass the threshold; the author checked the modules and added the AccessControl pattern.
Showed that one shared lock per backup let any account stretch or shorten a holder's stress lock, and that the backup key alone could drain the funds; the author confirmed both and will update the draft.
Showed that a single-use authorization kept in LangGraph graph state comes back unused when the graph is replayed from an earlier checkpoint; the poster agreed it cannot live in rewindable state.
Showed that the state-diff recovery tool accepted reverted blob transactions, which anyone can send with copied calldata and their own blob; the author confirmed it and added a receipt status check.
Showed that a node revoked at Tier 1 keeps the error vector it decoded earlier, and that two frames under one policy expose the XOR of their session keys; the author agreed on both.
Pointed out that anyone could list apps in a space, that maxToll gave users no tolerance for price changes, and that information-only edits voided pending operations; the author agreed on all three.
Pointed out that rateAfter is only a starting point on adaptive rate models, which keep rising above target utilization, and that ranking markets by rate ignores collateral; the author added the drift to the draft.
Pointed out that invoices keyed by id alone could be squatted, that mainnet USDT would make the payment call revert and that fee-on-transfer tokens could overpay refunds; the author fixed all three in the reference contract.
Showed that the opt-out marker permanently broke ERC-721 and ERC-1155 receiver checks and only held on one chain; the author made the EVM treat the marker as empty code and documented the per-chain scope.
Listed what a Hub price oracle must settle before contracts rely on it: price freshness, validator participation, thin-market listings and IBC relay delay; the author carried each point into the proposal's spec sections.
Suggested making operation type, target and the Safe transaction hash mandatory on the signing device and recomputing that hash away from the web interface; the author agreed and will make it a normative requirement.
Showed that a de-risk-only suspension on a parent binding blocked repay and cancel on derived bindings, and that a fulfiller could lift its own guardian's suspension; the author confirmed both and fixed them in revision 3.
Showed that a policy reinstalled after a switch away and back kept a stale holder count, which could admit holders past the cap or block a full exit; the author made reinstalls count as new installations and added tests.
Showed that the signed mining attestation named neither the receiving address nor the chain, so it could be resubmitted to another address or replayed on mainnet; the author confirmed it and moved to a v4 attestation covering both.
Showed that the payment check refused mints and self-transfers but not the owner's own top-ups, and that settlement contracts hide the real payer; the author fixed the check, added a named writer and documented settlement payers.
Pointed out that canonical bridges alias senders that have code, so deposits from a migrated account land on an alias address; the author confirmed it and added it to Residual risks with a wallet warning.
Showed that a fixed bond can be slashed for double voting but not for stalling finality, and that the 2-of-3 committee stays the real trust boundary; the author agreed and will split admission and upgrade keys with an upgrade delay.
Showed that earlier unlock events would still show tokens as unlocked to event indexers after a permanent lock, and raised proxy upgrades and fully burned ERC-1155 ids; the author fixed all three in the PR.
Pointed out that a lender reporting from two addresses could count one default twice and that a mistaken record could never be voided; the author will add reporter primaries, a void window and an owner requirement.
Pointed out that the claim NFT and the note secret could drift apart, that proofs need the recipient bound in, and that odd amounts link deposits to claims; the author agreed on all four points and bound the NFT into redeem.
Suggested a shielded refund note for IBC timeouts, so refunds do not expose a transparent address, and a per-denom balance counter in the gateway module to cap a circuit bug; the author agreed with both.
Showed in the testnet contracts that a leaked oracle key could still backfill days before an installation existed and that renouncing ownership would lock the founder vesting forever; the author fixed both and redeployed on Celo Sepolia.
Suggested sizing the default signing tree from the batch rate, a committee size rule that covers frozen keys, and asked how key reuse across guards is caught; the author agreed and said the ERC will state these rules.
Pointed out that for a small CPU-mined chain a deposit reorged away after minting on BNB Chain is a bigger risk than the multisig; the author agreed and planned deep confirmations, a daily mint cap and a pause switch.
Showed that anyone could call invest() to move tax reserves into the yield strategy without checking due dates, so a fully lent pool could hold up a tax payment; the author confirmed and now keeps reserves due within 14 days liquid.
Suggested rejecting float opcodes at deploy, canonicalizing NaNs in the engine and keeping relaxed SIMD off so untrusted wasm runs the same on every validator; the author said they will reject float opcodes at deploy.
Suggested putting a client-generated order id or nonce in the signed order, so a retry from a second tab can never be filled twice; the author agreed the guarantee belongs in the order and said they will look at it.
Showed that the pay page checked the merchant's protection policy and then sent a separate plain transfer, so a policy switched off mid-checkout paid the merchant unprotected; the author made the check and payment one atomic step.
Pointed out that an approve() call passes an ABI check while granting an allowance to the wrong spender; the author added approval decoding and an expected-spender check.
Suggested probing MCP servers with a valid token issued for a different resource, which the spec says they must reject; the author is adding that probe to the scanner.
Pointed out that a seller can farm repeat buyers for cents, so a trust score should trace where buyer wallets got their USDC; the author said the tool will add funding-wallet tracing.
Pointed out that a SQL pre-check tool is only advisory, since the agent can check one query and run another; the author agreed and added a mode where the server runs the query itself and the agent never holds database credentials.
Suggested pinning the token and network next to the recipient in an x402 spending sidecar for agents; the author shipped those pins and now stores them in each approval.
Showed that if another transaction took a payment's nonce, the payment would stay submitted forever and stall every payment queued behind it; the author agreed it was not handled and plans to detect the taken nonce and reconcile.
Proposed a test where two sibling blocks spend the same shielded note in parallel; the author added it as a unit test, then rebuilt it end to end in both block orders with a negative control.
Suggested replaying a valid x402 payment against a pricier route and swapping the token or network; the author added mint and network substitution checks to the suite.
Suggested reject vectors for the leader VRF's own point decoding, such as y >= p, x = 0 with the sign bit set and small-order points; the author added them and checked them against an independent decoder.
Suggested that the quote API decode the calldata it hands to agents and check router, recipient and minimum output before they sign; the author shipped that check and later verified Relay orders the same way.
Code we got merged into other projects, and our own open-source builds, are on smart contract development →
The checks behind these reviews, applied to your contracts, protocol design, agent wallet or MCP server before it ships.