Fund Safety Comparison: Custody Structures of the Leading Perp DEXs
On Oct 29, 2026, $6.3B left Binance's hot wallets: its own signing system approved an EIP-7702 delegation handing wallet control to the attacker's contract. A month earlier, Bitget lost $350M+ the same way — a backend feeding transactions to signers. The common lesson: the attack surface was not onchain; it was the signing and backend layer. An external researcher, @0xJaehaerys, used this lens to compare fund safety at four leading perp DEXs (original thread, Oct 31, 2026). This page reproduces that comparison with our design mapped against it; structural figures were independently double-checked (Nov 3, 2026), and TVL/balance numbers are as of the original post.
The three questions:
- Where does the money actually sit? (contracts, a custodian, or the project's own addresses)
- Whose signature moves it? (the user, a multisig, or a single operator key)
- If the team is hacked or vanishes, can you still leave on your own? (an exit path that needs no project cooperation)
Comparison
| Venue | Where the money sits | Whose signature moves it | Self-exit? | Bar to change the rules |
|---|---|---|---|---|
| Lighter (~$1B) | A single Ethereum contract; state changes verified by ZK proofs | Only proof-verified batches; standard withdrawals go only to the wallet linked at first deposit; sequencer/prover keys cannot pull balances | ✅ Forced withdrawal unprocessed for 14 days triggers "desert mode" — everyone exits from onchain data | Regular upgrades: 3-of-5 multisig + 21-day public notice; a 4-of-7 security council co-sign bypasses notice (every upgrade so far took that path) |
| Hyperliquid ($6.5B USDC) | ~86% in a single plain Coinbase address (visible onchain); ~10% a Circle contract; ~4% an Arbitrum bridge | Coinbase's key and the chain's own system transactions both move it; bridge is 3-of-4 hot keys | ❌ No self-exit | Last brake on the USDC side is the stablecoin issuer freezing; foundation validators hold heavy governance weight |
| Aster (~$500M) | Pooled vaults on four chains; ownership recorded on the project's own chain | Every withdrawal is sent by an operator address and co-signed by two operator keys; the user's signature is only checked against an address the operator supplies | ❌ No self-withdraw function in the verified code | 4-of-7 multisig + 6h timelock — but the same multisig can instantly change caps and roles; the Solana program is upgraded on the spot by 3-of-6 |
| Variational ($290M) | ~208k per-user settlement pools on Arbitrum | Any one of 17 operator keys (16 plain wallets) can move any pool, with no cap | ❌ No exit if the backend stops signing | A single admin plain wallet, effective instantly |
| Catswap (design target, not yet live) | Onchain program-owned vaults (PDAs); USDC held directly onchain | User-signed instructions to close and withdraw; terminal payout is a permissionless crank — anyone can drive it, the destination is pinned by the protocol to the beneficiary's account, and no operator withdrawal path exists at the protocol layer | By design, yes: payout needs neither the beneficiary's signature nor a live backend | Pre-launch requirement: upgrade-authority multisig + timelock (this page updates once finalized) |
Notes per venue
- Lighter: the only one of the four where hacking the servers still cannot pull your balance — trust collapses onto price feeds and multisig speed. Its named weakness is precisely price: index/mark prices are operator-supplied and the proof system does not verify price signatures, so wrong-price liquidations count as a fund-safety issue too.
- Hyperliquid: the largest, and the most concentrated — 86% of the USDC in one plain address. The JELLY incident (validators voted to delist and settle at their chosen price) shows its validator set can intervene in position settlement.
- Aster: the user's signature only proves "you asked", not protection for the pool; the hourly withdrawal caps ($20M–$30M per chain) have been hit repeatedly and lifted instantly by the same multisig.
- Variational: the most exposed on paper — a single operator key can drain any user pool with no cap, a batch function pays out to the caller's own key, and the address documented as the "vault contract" has no code.
Our design mapping
The hardest of the three questions is the third (self-exit), and the fourth column — the bar to change the rules — decides how long the other answers hold: every onchain contract guarantee is a guarantee under the upgrade authority. Three principles follow:
- Upgrade-authority governance: multisig + timelock + published upgrade notices, with freezing available in the endgame — withdrawal constraints must not fail at a single party's will.
- Allow-list signing: backends and executors pass only known instruction shapes instead of blocking known-bad ones — the shared root cause of both incidents (attribution is the researcher's inference; the parties involved have not confirmed it).
- Price integrity: median of multiple sources plus a cross-check gate, so the mark price is never decided by a single party — exactly the item Lighter was flagged on.
Practical advice for users (from the original thread)
- Keep only the funds you actively trade on any venue.
- Before parking size, check two things: where the money actually sits, and whether you can leave without the team.
Sources and caveats
- Original thread: Binance got drained. Would your Perp DEX? (@0xJaehaerys, Oct 31, 2026; evidence = onchain reads and verified contract code; the author states it is not an audit, offchain MPC and policies are invisible, and nothing is investment advice).
- Structural figures on this page were independently double-checked (Nov 3, 2026).
- The Catswap row describes design targets, not a live deployment; a pre-launch security disclosure will be published separately.