Hook
The first thing that caught me wasn't the $1.28 million. It was the 98 seconds.
A single wallet opened and closed 305 positions on Papertrade, a derivatives venue that had been live for roughly eight hours. Two hundred ninety-six of those trades printed green. That's a 97% hit rate, on a platform nobody had stress-tested, with a median holding time of ninety-eight seconds. I've run trading bots long enough to know what that distribution actually is. No human holds a book for exactly ninety-eight seconds, three hundred times, and wins 97% of them. That's a script with a cron job and a very specific idea about where the price was going.
The second thing was the candle. During the largest leg of the operation, ETH walked from 2507.35 down to 2504.4 — a move of $2.95, or 0.12% — on roughly $16.5 million of sell pressure. Inside that 0.12%, a $297 million notional short position on Papertrade was liquidated, and $1.28 million of somebody else's money moved into somebody else's wallet.
I've audited oracle integrations before. This one had every fingerprint of a price feed that was chosen for speed instead of safety.
Context
Papertrade is a perpetuals platform that never built its own price discovery. It borrowed Hyperliquid's — specifically the BBO, the best bid and offer on Hyperliquid's order book, rather than Hyperliquid's official oracle. That distinction is the entire story.
A BBO is a mirror. It reflects the instantaneous state of a single venue's book, and any aggressive order of sufficient size bends it immediately. An official oracle does something structurally different: it aggregates multiple venues, weights them across time, and usually carries a deviation band that refuses to publish a price that has jumped too far, too fast. One is a live wire. The other is a circuit breaker wrapped around a live wire. Papertrade wired itself directly to the first one.
If that sounds familiar, it should. In 2022, Mango Markets was drained for roughly $110 million by a trader who pushed the price of MNGO on thin books, then borrowed against the inflated collateral. Avraham Eisenberg argued publicly that it was a legal, profitable trading strategy. A jury disagreed, and he was convicted on fraud and market manipulation charges. The pattern here is the same shape at a different scale on a different venue.
The wallet behind the Papertrade trades isn't a tourist, either. It has run roughly $14 billion in cumulative volume on Hyperliquid and booked more than $13 million of profit there. Those are market-maker numbers. That's somebody with real infrastructure, real execution, and a precise understanding of where the price on each venue actually comes from.
The other detail worth sitting with is the timeline. Papertrade had been live for about eight hours. No audit gets done in eight hours. No adversarial testing happens in eight hours. Platforms that launch this fast are optimizing for a listing date, not for surviving the first whale who reads their docs — and in DeFi, the first whale who reads your docs is usually the one who empties them.
Core
Here's the mechanism, decomposed into its parts.
The attacker built opposing exposure on two venues at once. On Papertrade, where the borrowed price settles trades, they stacked roughly 30 short positions carrying about $297 million of notional value. On Hyperliquid, where they could actually move the price with real capital, they sold ETH to push the BBO down. Shorts profit when price falls, so the direction stayed consistent: push Hyperliquid lower, collect on Papertrade, where the position was an order of magnitude larger than the capital doing the pushing.
The capital efficiency is the whole game. Roughly $16.5 million of real ETH was used to try to move a $297 million position. You don't need to be right about ETH. You only need the venue you're harvesting to keep quoting a price you can bend. That's the trade, and it's a trade that exists purely because two systems disagree about what "the price" means.
Three hundred five trades, 296 winners, 97% hit rate, 98-second median holding time. That's not skill; that's automation. It's also exactly what an overfit reward function looks like when it finally meets a live book. I've built this thing. When I ran three arbitrage bots across OpenSea and LooksRare in 2021, gas fees ate 60% of a $50,000 principal, but the part that worked taught me the real lesson: the edge in cross-venue trading never lives in the signal. It lives in the plumbing. Here, the plumbing was Papertrade quoting a single order book it did not control and could not defend.
Then the strategy hit its ceiling.
The final round is where it broke. The attacker pushed about 6,600 ETH — roughly $16.5 million — of sell pressure into Hyperliquid and moved the price from 2507.35 to 2504.4. A $2.95 move. That was not enough to cover the profit requirement on a $297 million short book. Manipulation cost scales with how much liquidity you have to eat, while the payoff scales with position size — so once your position is larger than the book, you're just paying slippage to yourself. The price snapped back to 2509.95, the shorts were liquidated, and that single round cost roughly $450,000. Nine losers out of 305, and one of them was the tail that the 97% win rate was quietly hiding.
Strip the numbers back and you can see the prerequisite the whole operation depends on: a measurable gap between the price Papertrade quotes and the price the market is actually clearing at. That gap doesn't need to be large. It needs to be reliable. Scanning for that kind of dislocation is what my own bots were doing in 2021, just with fewer zeros on the end — hunting the same inefficiency, finding it in the spread between two venues rather than in a borrowed feed. The difference is that arbitrage closes a gap, while manipulation opens one and charges the venue for the privilege.
Now look at what's missing from the design. No TWAP. No multi-source aggregation. No deviation circuit breaker. The platform ran for eight hours and the same play worked repeatedly, which tells you the risk controls didn't exist rather than failed. From my audit work on Solend's oracle price feed — the integer overflow I disclosed for a $15,000 bounty — I learned that the vulnerability is almost never in the contract logic. It's in the assumption about where the number comes from.
There's a ghost in the machine here that's worth naming. The Papertrade book was never empty — it was full of prices that didn't belong to it. Every quote it published was a Hyperliquid quote, reflected through a mirror that anyone with a large enough order could tilt. The venue was hedging its entire book with a number it didn't own.
There's one more layer worth pulling apart. That $1.28 million wasn't created. It was transferred. On a perp venue, the counterparty to a winning position is the liquidity pool, the market maker, or another user — so the profit landed on Papertrade's own balance sheet as a hole. And the leverage made it worse: 30 positions worth $297 million got wiped by a price move of roughly 0.1%, which tells you the platform allowed either extreme leverage or razor-thin margin. Both amplify the attack and the liquidation risk at the same time. If Papertrade ever issues a token, this is the event that reprices it.
Contrarian
Calling this an "exploit" is the lazy read, and it's the read that lets everyone else off the hook.
From the attacker's chair, this was arbitrage. A venue published a price, offered to settle trades at that price, and the price didn't match the market. Taking that offer isn't a hack; it's a withdrawal. Arbitrage is just patience wearing a speed suit — except this wallet wasn't patient, it was fast, and the speed suit fit perfectly. Every bug is a bounty waiting for the right eyes — the only thing separating a bounty from a heist is who ends up holding the invoice. Here, Papertrade's LPs, its treasury, or whoever sat on the other side of those 30 shorts are the ones who paid it.
The deeper blind spot is that this industry keeps treating "we use a live oracle" as a security feature. It isn't. It's an outsourcing of your attack surface. We've spent years pretending on-chain price mechanisms are emergent — that Aave's interest curve is "supply and demand" when it's a governance parameter wearing a lab coat, that a DEX quote is a market when it's a formula. Papertrade's mistake was believing the mirror was the market. When you borrow someone else's price, you also borrow everyone who can move it.
And notice the pattern nobody wants to say out loud. The real difference between competing deployment stacks was never technical. It's who convinces more projects to deploy first — and whoever deploys first inherits the risk first. Papertrade launched fast by renting a price. It inherited a manipulation surface it never audited and never owned.
Takeaway
So here's what I'm actually watching.
Any new perp front-end whose documentation says it quotes Hyperliquid's BBO rather than its oracle is now a red flag, not a feature. That one string is the whole audit. On ETH, the manipulated band is 2504–2510, and the liquidation cluster at 2509.95 tells you exactly where the leverage is stacked. In a bear market, thin books plus high leverage means this same trade is sitting available on any venue that borrows a price it can't defend — and more of them launch every week.
The question isn't how Papertrade got hit in eight hours. It's how many other platforms are quoting somebody else's order book right now, unpatched, and calling it security. When the algorithm breaks, we become the hedge — and in this market, most of us are hedging against the venue, not the token.
