Threat model
The attacks ZeroTrace is designed against, what stops each one, what risk remains, and the limits the design accepts.
Inside the vault, proofs hide who pays whom and how much. Entries and exits stay public: purchases, shields, sales and unshields show an address and an amount, and there protection comes from the number of users. For each attacker: what they try, what stops them, and what is left.
Attackers and defences
Chain analyst
- Tries: to link a purchase or shield to a later sale or unshield by amount and timing.
- Stopped by: only partly. Spends reveal nullifiers and new commitments, not which notes were spent. Splitting amounts, waiting and selling to fresh addresses help. See Staying private.
- Residual risk: high while the privacy set is small. A distinctive amount shielded and taken out soon after is easy to match.
Malicious bundler
- Tries: to change the recipient, the gas fee or the sale terms of an operation it relays.
- Stopped by: the proof. Recipient, caller, gas fee, sale terms and ciphertexts are hashed into a public input
that your
Spendsignature also covers; changing any field invalidates the proof. A minimumcallGasLimitin the proof stops the bundler starving a sale's execution. - Residual risk: none on content. It can still see your IP address and timing.
Censoring bundler
- Tries: to refuse or delay your operation.
- Stopped by: any other ERC-4337 bundler, or a direct call. A wallet can submit through the EntryPoint's
handleOpsfrom your own address, and the vault'stransactcan be called directly. - Residual risk: a direct call reveals the address that pays gas.
Spammer draining the gas deposit
- Tries: to make the vault's EntryPoint deposit pay for operations that pay nothing.
- Stopped by: the private state change happens in validation, so execution cannot undo it. An invalid proof writes nothing and is not executed. A valid operation pays the network fee, worth the operation cost at the price average (0.0002 ETH at the start), and its maximum gas cost may not exceed that same operation cost, so it never takes more from the deposit than it pays. When the guardian changes the operation cost, an operation that still pays the replaced fee is held to the replaced cost.
- Residual risk: the fee is valued at the price average, which trails the market and is converted into ZERO once an hour. After a fast drop in ZERO's price the fee is worth less than the operation cost until the average catches up, and holders bear the difference.
Price manipulator
- Tries: to move the vault's price average so the network fee is underpriced, or so the harvester buys dear.
- Stopped by: the average reads only ZERO's own Pons venue: the curve's reserves before graduation, the Pons
pool's
slot0after. Only the price at a block's last poke counts, folded in when a later block is first poked, with weight 1/8 and clamped to ±10% of the average: at most 1.25% per block (an L1 block, about 12 seconds, on Robinhood Chain). The network fee follows the average only once an hour. The harvester buys at most once per L1 block, each buy moving the price by at most 0.5% and paying no more than 15% above the average after Pons fees; a buy that does not fit is halved, over up to 8 sizes, then left for a later harvest. - Residual risk: anyone can poke, so a price pushed, poked and pushed back in one transaction still counts for that block, within the clamp. Repeated over many consecutive L1 blocks, each paying the Pons fees on both legs of the push, and held until the hourly fee update, it can make the network fee cheaper for an hour or two. Each operation still draws at most the operation cost from the deposit. See the accepted risks below.
Decoy pool
- Tries: to open another ETH/ZERO pool, at an absurd price, so the router trades there or the vault prices from it.
- Stopped by: the router, the harvester and the vault use only the pool key Pons graduates into (native ETH,
ZERO, the launch's fee and tick spacing,
factory.memeHook()). Only the Pons factory can initialize a pool with that hook, so nobody can open that exact pool first. Any other key has another pool id. A fork test opens a decoy and checks it is ignored. - Residual risk: none for ZeroTrace. Third-party routers may still find the decoy.
Sandwicher
- Tries: to trade before and after your buy or sale, or around a harvest.
- Stopped by: minimum outputs from the router's
quote(Pons fees included) less 1% slippage. For sales the minimum ETH out is bound to the proof. Harvest buys are at most one per L1 block and move the price by at most 0.5%: pushing the price up before one and selling after costs the Pons fees on both legs, about 2%, for at most the 0.5% it moves. - Residual risk: up to the slippage you accept. Robinhood Chain orders transactions by arrival, which makes
sandwiches harder but does not rule them out. The harvester sweeps fees with no minimum outputs
(
sweepFees(0)andsweepPoolFees(poolId, 0, 0)), so a conversion of fees that Pons performs inside that sweep has no bound set by ZeroTrace.
Launch sniper
- Tries: to buy ZERO in the first moments after launch.
- Stopped by: Pons' opening snipe tax on buys, 99% decaying to 0 over the first 5 seconds. The deploy script buys the seed only once the tax is over.
- Residual risk: buying after the first 5 seconds is open to everyone, by design.
Pons redirecting the fees
- Tries: the Pons owner calls
setCreatorFeeRecipientfor ZERO, or uses its rescue functions on the curve or the graduation (rescueCurveFees,forceSweptGraduation,rescueSweptGraduation). - Stopped by: a 3-day timelock on the fee redirect, visible in
pendingCreatorFeeRecipient. ZERO has no owner and no transfer restrictions, so nothing Pons controls can move the vault's ZERO. - Residual risk: holder yield stops, and the gas deposit is no longer topped up with fee ETH; anyone can still add to it. Trading and the price average depend on Pons' curve and pool. Notes stay spendable: sends and unshields never touch a Pons contract. See Trust model.
The team's keys
- Tries: to misuse the vault's roles or the harvester's ownership, all held by the deployer at launch. The upgrader replaces the vault's implementation; the guardian pauses every spend, pauses entries or exits handed to a target, or raises the operation cost; the admin grants any role to any address; the harvester's owner hands ZERO's creator fees to a contract that keeps them.
- Stopped by: for the guardian, the limits of its functions: a pause moves no note and no ZERO, and the
operation cost stays between 0.00005 and 0.001 ETH. Every action is public:
Paused,EntriesPaused,TargetExitsPaused,OpCostSet,RoleGranted,RoleRevokedandUpgradedon the vault,FeeRecipientHandedOveron the harvester. - Residual risk: the upgrader can install any implementation, with no timelock, and that implementation decides what happens to every note and to the ZERO behind them. The guardian can stop every spend for as long as it keeps the vault paused, and nobody can take ZERO out meanwhile. The network fee can be raised to what 0.001 ETH is worth. The harvester's owner can redirect the creator fees at once, so the holder yield depends on the owner until ownership is renounced. See Trust model.
Keeper absence
- Tries: nothing. The risk is that nobody calls
harvestand the deposit runs dry. - Stopped by:
harvestis open to anyone, refunds the gas of a call that did work, and a ZeroTrace keeper calls it. Anyone can also top up the deposit directly. With an empty deposit, or when gas costs more than an operation may draw from it, calling the vault'stransactdirectly (an unshield) or the router'ssellstill works while the vault is not paused, with gas paid by the caller. - Residual risk: fees wait in the Pons escrow until someone harvests, and a direct call reveals the address that pays gas.
Tree exhaustion
- Tries: to fill the note tree (depth 24, about 16.7 million leaves) so no new notes fit.
- Stopped by: a full tree rolls over to a new one. Its final root stays valid forever, so its notes stay spendable. Each proof uses one root, so both inputs of a spend come from the same tree.
- Residual risk: filling a tree costs gas for every insertion.
Lying indexer
- Tries: to hide or alter notes, or to link users to their notes.
- Stopped by: the same event ranges for everyone, and a root check with
isKnownRoot; on a mismatch, or when the indexer is down, the app rereads the chain. See The Merkle tree. - Residual risk: the root check covers commitments only. An indexer that altered a note's ciphertext, or dropped
NullifierSpentorShieldedevents, would pass it: the app would miss a note or show a spent one. Funds are not at risk, and reading from the chain gives the right answer. The indexer also sees IP and timing.
Phishing site asking for the activation signature
- Tries: to get you to sign the activation message on another domain.
- Stopped by: the message is a Sign-In with Ethereum message that names the app's domain, so wallets warn when another site asks for it. A stolen activation signature reveals your notes but cannot spend them: every spend needs a fresh EIP-712 signature from your wallet.
- Residual risk: your privacy is lost for good. The keys are deterministic and cannot be rotated for that wallet; moving your notes to a new wallet restores it for future activity. A site that you then use to sign spends can also get you to sign a harmful transaction (see Trust model).
Intercepted payment link
- Tries: to claim a payment link meant for someone else, by reading it in the message that carried it, a shared screen or a browser history.
- Stopped by: nothing on-chain: a link is a bearer key, and the first claim wins. The key sits in the URL fragment, which browsers never send to a server, so the site and link previews do not see it.
- Residual risk: the amount of that link. Share links only with the person you pay, over an end-to-end encrypted channel, and take back links that stay unclaimed.
Bug in the circuits or contracts
- Tries: to exploit a flaw, for example to create shares from nothing or spend a note twice.
- Stopped by: tests of the contracts and the circuit. The vault totals B and S are public, and every change emits
VaultUpdated, so anyone can check them against the vault's ZERO balance. The guardian can pause every spend, and the upgrader can replace the vault's implementation, verifier included, keeping every note. - Residual risk: the vault has no cap on the ZERO it holds. A pause or an upgrade protects what is still in the vault. The router and the harvester cannot be upgraded; a new one takes a new deployment.
Stated limits
- Small privacy set early on. While holders are few, "who" is weakly protected. See Staying private.
- Private by choice. ZERO held outside the vault is an ordinary public balance. Only shielded ZERO is private, and only notes earn the holder yield.
- Public entries. A purchase with public ETH, or a shield of public ZERO, shows the address and the amount. What stays hidden is which note is theirs and when it is spent.
- IP and timing metadata. The RPC node, the bundler and the indexer can see IP addresses and when requests arrive. The app scans for notes only on demand, not on a schedule, so a scan right before a spend can hint at it (the wallet page does read the public network figures every 15 seconds).
- Dependence on Pons. Trading, the price average and the holder yield run on Pons contracts and on Pons' fee payments. See Trust model.
- Roles held by the team. The vault's upgrader can replace its implementation with no timelock, its guardian can pause every spend with no time limit, and its admin can grant either role to anyone. The harvester's owner can hand the creator fees to another contract at once. At launch the deployer holds all of them. See Trust model.
- Contract-based multisigs are unsupported. The circuit checks one EOA ECDSA signature.
Accepted risks
These behaviours are known and accepted in the current code.
| Behaviour | Effect | Why accepted |
|---|---|---|
claimSale can run in the same bundle, before a sale executes | The sale's ZERO goes to the ETH recipient instead of being swapped | The attacker gains nothing. The EntryPoint gives no "same transaction" signal, and ERC-7562 bans timestamp and block reads in validation |
| The network fee is flat, worth the operation cost in ZERO (0.0002 ETH at the start, set by the guardian between 0.00005 and 0.001 ETH) | An operation pays the same fee whatever gas it uses, and the fee goes to the other notes. When gas costs more than the operation cost, the vault refuses ERC-4337 operations | The fee is known before proving and cannot fail because gas or the price moved while signing; the deposit never pays more for an operation than its fee is worth. Direct exits, paid by the caller, remain while the vault is not paused |
| A harvest buy that does not fit the 0.5% impact and 15% premium bounds waits | Fee ETH waits in the harvester for a later harvest | Buying in larger steps, or far above the average, would give value away to whoever moved the price |
| A block's price is the price at its last poke, and anyone can poke | A price pushed, poked and pushed back in one transaction enters the average, clamped to ±10% and with weight 1/8 | Each block costs the Pons fees on both legs of a push large enough to move the price 10%, over many consecutive L1 blocks. The most it can win is a cheaper network fee for an hour or two, each operation still drawing at most the operation cost, and the harvester stops buying once the market is more than 15% above the average |
| A buy that crosses graduation can be partly refunded by the curve | The router returns the unspent ETH to the buyer; the harvester keeps it for its next harvest | The curve stops selling once it is full; the rest trades on the pool after graduation |
Trust model
What you have to trust when you use ZeroTrace, what you do not, and what each party, Pons included, could and could not do to you.
Post-mortem, 12 October 2026
How the earlier ZeroTrace vault was emptied by a forged proof that a flawed Barretenberg verifier accepted, what we have done, and the refund of every note.