Risks and privacy
What can go wrong for users and for the vault, what limits it, and what a position reveals.
The vault guarantees a few things regardless of market conditions:
- users are isolated from one another, and reserved USDC is protected;
- exits always work, because closes, liquidations, added margin, and claims are never blocked by a cap or a pause;
- every protocol change waits one day;
- all exposure is capped relative to protocol equity.
The risks below are what remains.
What users see
| Risk | What the user sees | What to show |
|---|---|---|
| Price moved past their limit | The request is cancelled with a full refund and no fee | The reason, and a retry |
| Request not executed | A full refund after two blocks | Retry automatically |
| Liquidation | The position is closed; the penalty goes to the liquidator and the buffer, and any leftover is returned | The liquidation price and health |
| Opens blocked by caps | The request is cancelled at execution | The cancel reason |
| Profit delayed | Part of claimable waits for USDC to be bridged back from HyperCore | The claimable amount, and a retry |
MIN_HOLD_TIME | A close is rejected for 10 seconds after opening, unless liquidatable | A timer |
| Paused | No new opens; everything else works | A status banner |
Instant fills
Between a fill and the hedge filling on HyperCore, one to three blocks, the
vault holds users' net exposure itself. In normal markets the move over a few
seconds is below spread plus fee. In a crash it can be several percent. The loss
is bounded by about maxUnhedgedUsdc times the move during the lag, because
risk-increasing opens are cancelled once the gap would exceed the cap.
A trader who sees a move elsewhere before Hyperliquid's mark updates could try
to trade against the old price. Next-block execution means the fill price is
unknown at request time, the two-block window stops a requester from waiting
for a favorable move, and the spread, trade fee, MIN_HOLD_TIME, and unhedged
cap all limit the edge. A small edge remains if Hyperliquid's mark itself lags
the wider market by more than a block.
Hyperliquid's taker fee and slippage on hedge orders can cost more than spread
plus fee earns; netting across users reduces hedge volume. A hedge IOC can also
miss in a fast or thin market. Nothing rests on the book, the next hedge()
retries from szi, and the unhedged cap stops new risk while the gap is large.
If nobody calls execute within two blocks, the request expires and is refunded
in full from block N+3. Anyone can execute, including a user's own relayer.
Market
A price can gap through a liquidation level before liquidate runs, and losses
can exceed the insurance buffer. Leverage tiers tighten as open interest grows,
liquidation is synchronous at the current mark, liquidators are paid even for
underwater positions, the penalty refills the buffer, and gross open interest is
capped relative to protocol equity.
The vault's HyperCore position runs at about 3x leverage. If nobody tops up
margin in a fast move, HyperCore can liquidate the hedge. fundCorePerp is
permissionless, and if margin runs short, hedge orders stop filling and the
unhedged cap stops new opens.
The hedge pays funding and ties up margin. The borrow fee charges every position for this, pro rata on entry notional. A rate change applies only to time after it, so raising the rate cannot make a position liquidatable for time already held.
An attacker could push Hyperliquid's mark on a thin market to trigger
liquidations. Next-block pricing, the caps, the leverage tiers, and the
immutable LEVERAGE_CEILING limit this. Thin markets should be deployed with a
low ceiling and small caps.
Settlement
Close payouts are credited on HyperEVM, but hedge gains accrue on HyperCore.
claim() pays what free HyperEVM USDC covers and leaves the rest claimable
until bridgeSpotUsdcToEvm brings funds back. Users see a delay, not a loss.
Bridge delays make this last longer.
The vault credits claimable and never calls into a clone, so a payout takes
one extra call. In exchange, a hostile clone cannot block or re-enter a payout.
Operations and platform
An attacker holding the protocol multisig keys can pause opens and propose bad
parameters or an insurance withdrawal. Every change waits one day and must pass
the on-chain parameter bounds, and a withdrawal must keep the equity floor. The
protocol key should be a multisig, and every propose event should be
monitored.
If the protocol's bot stops, anyone can execute, hedge, and liquidate, requests
expire with a full refund, and liquidators are paid. Callers choosing when to
call hedge, rebalanceCoreMargin, or execute are limited by the hedge
cooldown, the dust and bridging minimums, the rebalance band, the once-per-block
margin rule, and the two-block window.
The vault depends on HyperCore's mark, position, and account summary precompiles and on CoreWriter's IOC order action. If HyperCore stops, hedging stops, the unhedged cap blocks new risk, and exits still work at the mark. If the mark itself stops updating, the protocol should pause.
The vault has not been audited. It calls only USDC, CoreWriter, and the read
precompiles; every state-changing entry point is nonReentrant and writes
state before token transfers. Margin rounds up and payouts round down, and
packed storage is written through SafeCast, so an out-of-range value reverts
instead of being truncated.
Privacy
A position is opened by a fresh, single-use clone funded by an unshield, so the vault never sees the user's address. HyperCore sees only the vault's net IOC orders across all users. Some information still leaks:
| Leak | Mitigation |
|---|---|
| Amount and timing link an unshield to a request, and a payout to a reshield | Standard denominations and optional random delays |
| Position size and leverage act as a fingerprint | Common sizes and leverage |
| Small anonymity set at launch | Grows with use |
| The gas payer | A relayer paying gas from the note |
Use a fresh clone every time
Reusing a clone salt across positions links them. The vault already rejects a second open from the same clone, so each position needs a new salt.
Continue with the privacy model for the pool's own guarantees.