For agents and builders
Audits
Two audits of Toll by Toll, both on 17 Nov 2026: one of the money paths and secrets, one of how often the checks are wrong. Every finding, what was done about it, and what is still open.
On this page
Whose audits
These are Toll’s own audits, not an outside firm’s. No third-party security firm has audited Toll yet. We publish ours so anyone can see what was looked at, what was found and what was done. Findings are described as the kind of problem and the fix; no secret, internal address or step to reproduce is published.
The rules themselves are open on Rules & thresholds, and every pre-sign answer is scored 24 hours later on the pre-sign record.
Security: money paths and secrets
Method. A read of everything new or changed since the 6 Nov 2026 audit and the 11 Nov facilitator review: the x402 facilitator’s settle webhooks and seller dashboard, the public payee re-check, ALLOW tickets, paid jobs, the Telegram desk shop and the token text that reaches agents. Read-only: no transaction, no state-changing call, no deploy; live checks were reads only, and server settings were read by name only.
Result. 0 critical, 0 high, 3 medium, 6 low. Eight are fixed; one is open on a feature that is switched off.
- M1 (medium, fixed). A seller's settle webhooks shared one queue with everyone else's, and refused-payment events had no per-seller limit, so one party could crowd out other sellers' events. Done: Refused events are now limited per seller before they are queued, the same limit as the refusal record itself.
- M2 (medium, fixed). The public re-check of a seller's payee could store a worse result from a URL the caller chose, so anyone could make a verified seller's public record look unverified. Done: A URL the caller names can now only add a verified result; it can never replace the record with a worse one.
- M3 (medium, fixed). An ALLOW ticket could be signed as a WARN for a swap whose coin was never read (busy, or past the free daily coins). Done: No ticket is signed while any coin in the transaction was not read.
- L1 (low, fixed). A paying party could fill another seller's refused-payment list and webhook within the rate limits. Done: Covered by the M1 limit; the code comment that said this could not happen was corrected.
- L2 (low, fixed). After an agent changed hands, the old owner kept write access to its seller dashboard for a few minutes, and a changed webhook could keep receiving events briefly. Done: Writes now read the owner live, and the webhook cache was cut to 10 seconds.
- L3 (low, fixed). The public re-check could be used to make Toll fetch outside URLs repeatedly. Done: Re-checks are now also limited per target host and per caller.
- L4 (low, open). In the Telegram desk shop, the address check of a seller's host was not pinned to the connection that follows. Done: Open. The desk shop is switched off and cannot sign anything; this is fixed before it is ever turned on.
- L5 (low, fixed). A wallet clean-up or bot replay could be bought for someone else's wallet and announced publicly by our bot account. Done: The bot now posts about a wallet only when that wallet paid; the stamp refuses other payers outright.
- L6 (low, fixed). Token names from one experimental launch source reached agent-facing text with only basic control characters removed. Done: They now go through the same cleaning and AI-instruction scan as every other token text.
Checked and found sound, among others: webhook signing and replay window, webhook address checks, how the dashboard proves a seller, that client-sent IP headers are not trusted, the facilitator’s caps under concurrent settles, ALLOW tickets (never on BLOCK, bound to one exact call on chain 4663, single use, 60 seconds), and the desk shop’s spending caps.
Accuracy: how often the checks are wrong
Method. Live answers from the site and the pre_buy_check tool compared with what happened on chain, read from the public RPC (each pool’s own liquidity history and transaction receipts), and with the launch record’s 24-hour outcomes from 11 to 17 Nov. “Rugged” is the scorer’s own outcome 24 hours after launch. The Hookr report was checked against Hookr’s published contracts, not against Toll’s own decoder.
- F1 (critical, fixed). "A wallet sold X% of supply into the pool" counted a liquidity add as a sale. A dev seeding its own pool read as selling 100%, and that is a BLOCK. 14 of 14 cases checked on chain were liquidity moves, not sales. The coins mostly rugged anyway, but the stated reason was false. Done: Liquidity adds and removes are no longer counted as sales: Uniswap v4 position-manager transactions are skipped, and v2 adds are told apart by the pair's Mint / Burn in the transaction receipt.
- F2 (high, wording fixed). The "bought in the pool's first seconds" rule BLOCKs at any share, but its wording said "a large share" even for 0.0%. Done: The wording no longer claims a size; the check page gives the share. The threshold was kept: the measured rate is for this launch pattern as a whole, and coins with it still rugged at about 96%.
- F3 (high, fixed). Coins deployed by a wallet instead of a launchpad, with only careful findings, rugged at 97.5% (39 of 40) yet could get ALLOW. Done: Such a coin now carries a WARN with the rate of every such coin (coin-nopad-base-rate). It is never a BLOCK: it is a base rate, not a finding on the coin.
- F4 (medium, fixed). When side pools were read but the main pool was skipped, the liquidity-pull scan counted as complete, so a pull in the main pool could go unseen. Done: A scan that missed the main pool now says "not read", and pre_buy_check answers WAIT.
- F5 (medium, fixed). The launcher-record BLOCK had no guard for a public launchpad crossing 80% rugged because of spam, which would have blocked its honest coins too. Done: Named launchpads never BLOCK on the launcher record; at that rate they get a careful note. Only an unnamed launcher contract can BLOCK.
- F6 (low, open). One scorecard field counted groups of calls instead of coins. Done: Open. No answer uses that field: the launcher record sums the per-coin counts.
- F7 (low, open). The first-seconds buyer share could read over 100% when buys and round trips were summed gross. Done: Open (data quality). The BLOCK on those coins was still right on the outcome.
- F8 (info, open). Signals the check does not have yet: a dev holding 5% or more from its launch buy, and the funder's own 7-day record. Done: Open: to be added and measured before either can affect an answer.
- F9 (medium, fixed). The weekly Hookr report left out some blocks a pool really has (LP rewards, creator royalty, launch windows). 13 of 210 fields checked against Hookr's published contracts; nothing listed was false. Done: Fixed: the guard that skipped those pools was removed, and the core blocks are now decoded from the pool's own config instead of logs that some blocks never emit.
Checked and found correct: a self-deployed coin’s base-rate line never becomes a BLOCK; 15 of 15 sampled self-deployed coins really were deployed by the wallet itself; a critical warning BLOCKs even when other parts are unread, and ALLOW needs every part read.
How often a BLOCK is wrong
Measured on chain in the accuracy audit:
- coin-launcher-record (the launcher's 7-day record): 0 of 30 coins checked on chain were wrong. every coin: one add by the dev, removed in full by the same wallet minutes later, pool now empty.
- coin-lp-pulled (liquidity already pulled): 0 of 33 coins checked on chain were wrong. the pool's own liquidity history: who removed it, and whether the pool is now empty.
- coin-strong-against (the calibrated findings): 80 of 2,432 coins did not rug (3.3%); 36 were still trading or graduated (1.5%). the launch record's 24 h outcome, 11 to 16 Nov.
- hook-sells-faill: not measured. no coin in the samples carried it.
The BLOCKs were almost never wrong about the outcome. Where they were wrong, it was about the reason (F1, F2 above), and both are dealt with. Limits: 0 of 30 is an upper bound of about 10%, and it holds for the two launchers that fire today; no launcher that does not rug has crossed the bar to test it on.
Still open
- L4: Open. The desk shop is switched off and cannot sign anything; this is fixed before it is ever turned on.
- F6: Open. No answer uses that field: the launcher record sums the per-coin counts.
- F7: Open (data quality). The BLOCK on those coins was still right on the outcome.
- F8: Open: to be added and measured before either can affect an answer.
- A third-party audit, by a firm with its own name on it.
- Moving Toll’s signing keys to a hardware or managed signer once volume justifies it.
What would make this stronger
An outside review. Until then: the rules are public and generated from the code, the measured rates are re-run every week, and every answer is scored after the fact. Found something we missed? Tell us on X (@tollchk) or in the Toll Booth channel.
Information only. Facts read from the chain, not advice.