Every Loss This Week Needed Someone Else's Key

Bitcoin printed 55,130 dollars and 1.8 billion dollars of leverage died. But every loss that will not come back happened where somebody else held the key: a bridge, a lending contract, a listless treasury company. The same week, central banks were flying gold home for exactly that reason.
Every Loss This Week Needed Someone Else's Key

Every Loss This Week Needed Someone Else’s Key

A vault door in a dark concrete room, its lock removed and a single key hanging on a nail beside it

A deadline nobody can enforce

On Monday at 18:38 UTC an offer expires that no court issued and no contract backs. Chainflip, a cross-chain swap protocol, has given the person who emptied its vaults until then to send back 90 percent of the money. In exchange it will call him a white hat, pay him a 10 percent bounty worth roughly $1.2 million, and withdraw its complaints to law enforcement.

There is no other lever. No insurance fund, no freeze function, no clawback, no committee that can reverse a signed transaction. There is an offer, a clock, and a paragraph addressed to the attacker noting that his addresses are flagged across every analytics firm and that every Tornado Cash withdrawal is being watched (TFTC, 22 November).

That is what a bridge is. This week made it literal.

The same week, bitcoin printed $55,130, its weakest level since September 2024, and then bounced 9 percent back above $60,000 (CoinDesk, 21 November). Trading near $59,700 this morning, it sits 52.7 percent below the $126,208 high of October 2025 — my own arithmetic on weekly candles read at 09:30 CET today.

Two numbers, one week. The first will be forgotten by December. The second will not.

Here is the claim I want to defend. Three different kinds of loss happened simultaneously, and only one of them is reversible. A price decline is a repricing: nothing left the building, the marks moved. A liquidation is a transfer between two parties who both chose the leverage. A custody failure is a change of ownership, and it does not come back when the chart recovers. Every permanent loss this week happened at the seam where somebody handed a key to somebody else — a validator set, a lending contract, an exchange. And in the same seven days, the world’s central banks were quietly flying gold home for precisely that reason, arriving at the same conclusion from the opposite end of the balance sheet.

What the price did, and what it did not do

Start with the selling, because it is well documented and none of it is mysterious.

About $1.8 billion of leveraged crypto positions were liquidated in 24 hours, $1.3 billion of them bets on higher prices (CoinGlass, cited by CoinDesk on 21 November). Glassnode had tracked a dense cluster of long liquidations between $59,000 and $62,000 for months; Friday cleared it, which the firm described as removing “a large source of forced selling” below the market. CryptoQuant’s Darkfost put seven-day realised losses above $11.3 billion, roughly eight times the gains booked over the same stretch.

11.3 billion dollars of realised losses in seven days, about eight times the gains booked over the same period

That last figure deserves a moment, because “realised” is doing the work. An unrealised loss is an opinion about the future expressed in today’s price. A realised loss is a signature. Somebody sold at a level they had promised themselves they would never sell at, and the loss became an accounting fact that no later rally edits.

What did not happen is just as informative. Open interest on Binance’s bitcoin perpetual market fell from 101,152 BTC on 13 November to 91,792 BTC this morning, a decline of 9.3 percent through a week in which the price fell 12.6 percent (own reading of Binance BTCUSDT daily open interest, 09:30 CET, 22 November). Funding is still positive: longs are paying shorts 0.0059 percent every eight hours as I write this. The dollar value of that open interest fell 23.5 percent, but almost all of that is the price, not the positions. The leverage was trimmed. It was not purged.

Open interest on Binance’s bitcoin perpetual market, in coins, at four daily snapshots in November 2026. Source: own reading, 22 November 2026, 09:30 CET

Bitcoin weekly closes from late August to 16 November 2026

This matters for what comes next, and it matters for the story people told themselves on Saturday. A capitulation that leaves funding positive and open interest within 9 percent of its peak is not the end of a credit cycle. It is a mark-to-market event with a lot of survivors still standing in the same posture.

The steel man: this was a rate event, and I should say so first

The strongest objection to everything above is that custody had nothing to do with any of it. Let me make that case properly, because it is a good one.

The two-year Treasury yield went from 4.24 percent in late August to 4.88 percent on 13 November, 64 basis points in eleven weeks (FRED, DGS2, observations through 19 November, when it stood at 4.76). The Nasdaq fell more than 4 percent on Friday and closed near bear-market territory. Bitcoin Magazine’s read on Friday was that bitcoin was trading “like a risk asset” with Fed hikes looming. When the discount rate moves like that, every long-duration asset gets repriced, and bitcoin trades with the longest duration of all because it produces no cash flow to discount.

US two-year Treasury yield, weekly, late August to 19 November 2026

On that reading, this week was a liquidity event. Bitcoin fell because the rate rose, crypto fell because bitcoin fell, and the specific failures, a bridge and a treasury company, were ordinary fraud-and-exploit noise that would have happened in a bull market too. Nothing about the drawdown required a theory of custody.

I accept the mechanism and reject the conclusion, and the difference is worth stating precisely. The rate explains the price. It does not explain the transfers. A 64-basis-point move in the two-year does not put $12 million into an attacker’s wallet, and it does not leave a Nasdaq-listed company 20 cents above a liquidation threshold that no one can renegotiate. Those two things happened because of who held which key. The rate decided that they would happen this week.

Twelve million dollars and a ten-day runway

Chainflip’s loss was first reported at $4.86 million on Saturday, when The Block and CoinDesk wrote up a vault drain. By Sunday the protocol had confirmed 122.86 BTC, 520.22 ETH, 3.25 million USDC and 758,141 USDT. At weekend prices that is about $12 million, which my own arithmetic on the four components adds up to within reporting error.

The stolen assets at weekend prices: bitcoin was six tenths of the loss

The mechanism is the part to read slowly. Chainflip’s validators collectively hold a threshold signing key that authorises movements out of its vaults. On 11 November, according to the company’s own account, the attacker funded preparation through Tornado Cash. On 20 November he registered a malicious validator account. On 21 November he used that position to compromise the aggregate signing key and moved the assets out. The transaction was, from the network’s point of view, perfectly valid: signed by the key that is supposed to sign.

Chainflip attack timeline, 11 to 23 November 2026

Note the runway. Ten days from funding to drain, and the protocol’s only defence was a validator onboarding process that let a new participant reach the signing threshold. Chainflip’s security architecture page had previously claimed that no single entity ever controls user funds during a swap. One compromised key, one malicious validator, $12 million in one signed transaction.

The company says it has identified the mechanism, taken steps to prevent a repeat, and holds identifying information on the validator and servers used. It has published no detailed post-mortem. That document is the single most important thing to read this week, and I will explain why when I get to the honest part.

FLIP, the protocol’s token, fell 30 percent within hours of the first reports and 42.9 percent over 24 hours by Sunday (CoinDesk, 21 November; TFTC citing CoinGecko, 22 November). The protocol entered “Full Safe Mode”, disabling deposits, withdrawals and swaps. There was no insurance fund, and no on-chain recovery path. The 122.86 BTC in that vault were real bitcoin, routed through a cross-chain bridge by people who wanted to swap. Every sat that takes that route is a bet on somebody else’s key ceremony.

This is the second Chainflip exploit of 2026; the September one hit a Tron vault and the funds were repaid. The difference in size is the difference between a bug and an architecture.

The loan nobody can renegotiate

Two days earlier, on the other side of the same lesson, a Nasdaq-listed company filed with the SEC and used a sentence that no banker would ever write into a credit agreement: “We can provide no assurance as to the future amount of our borrowings, the value of our collateral, or that a liquidation will not occur.”

BTCS holds ether as its main asset. It has borrowed about $54 million of dollar-pegged stablecoins against 46,676 ETH posted as collateral on Aave, a lending protocol that runs as code. It also carries about $18 million of senior secured convertible notes and held roughly $111 million of crypto valued at an ether price of $1,775 (Form 8-K filed 19 November, reported by CoinDesk on 21 November).

The debt was $64 million two weeks earlier. BTCS got it down by about $10 million in ten days, and CoinDesk reported on 13 November that the company had been selling ether below its purchase price to keep the position healthy. It did not help. Aave scores each loan with a health factor that turns a position liquidatable below 1.0. CoinDesk’s estimate put BTCS at roughly 1.24 on 9 November, with $64 million of debt and ether at $2,127. The smaller debt, with ether at $1,775, leaves the score at about 1.23; at Saturday’s price near $1,730 it is about 1.20. Ether bottomed at $1,536 on Saturday. The modelled liquidation level, assuming the collateral balance has not changed, sits near $1,450.

Ether daily closes, 9 to 22 November 2026, with the modelled liquidation level of about $1,450 for the ether-backed Aave loan of BTCS. Sources: Binance ETHUSDT daily candles, read 22 November 2026; CoinDesk, 21 November 2026

Here is the sentence in CoinDesk’s report that belongs in every discussion of “DeFi lending”: outside traders can repay part of the debt and take the collateral at a discount, with no margin call and no one to negotiate with.

Compare that to the bridge. Chainflip’s loss was a negotiation because a company stood behind it. BTCS’s position has no negotiator at all. If ether prints $1,450, the code executes, the collateral goes, and the shareholders discover that they had been running a levered ether fund with a stock price of 65.3 cents — down 12.2 percent on Friday and 28.1 percent on the week.

Both failures are the same failure. One is a key held by a validator set, the other is a key held by a smart contract whose parameters were chosen when ether was three times the price. In neither case did the holder of the asset control the moment of sale.

The same wall, underground

If you think this is a DeFi problem, read this week’s Bitcoin Optech newsletter, which summarised a Delving Bitcoin discussion about paying miners directly from the coinbase transaction. The proposal family is called decentralized coinbase payout protocols. The attraction is obvious: instead of a pool operator holding everyone’s reward until it decides to pay out, each miner gets an output.

Then the wall appears. Mining firmware limits coinbase transaction size, and every payout output displaces fee-paying transactions, so you cannot give every miner its own output. When a pool has more miners than outputs, any rule that ranks miners by work allows a rich-list attack: a large miner splits its hashrate across identities that each still outrank the smallest miners and pushes them out of the block. LearnBitcoin.com examined 2,545 P2Pool coinbase transactions and argued that growth alone, with no attacker at all, squeezed out small miners — P2Pool raised the share difficulty of miners expected to earn under 0.001 BTC per block.

And the fallback is the part that belongs in this article. For a single coinbase output paying a joinpool, miners vary the coinbase transaction for extra nonce space, so its txid is unknown until a block is found and exit transactions cannot be pre-signed. Until a soft fork adds rebindable signatures, that output is controlled by a single key or a threshold of keys. The same structure, one layer down: the thing you have to trust is whoever holds the key while the block is being found.

The same newsletter carried an Eclair disclosure with the same shape. A missing lower-bound check on a batch-size field let any peer crash a Lightning node without even having a channel with it. The specification says strictly greater than 1 and at most 20; the implementation only enforced the upper bound. No direct access to funds. But a node that can be crashed repeatedly is a node whose channels are at risk, and the fix shipped in July 2026 and the disclosure came four months later.

None of this is scandal. It is what an honest accounting of trust looks like: a list of the places where a key, a check, or a person stands between you and your money.

The other end of the balance sheet

Now the part that makes this more than a crypto story.

On 19 November the Official Monetary and Financial Institutions Forum published a report called Holding Power, built on four confidential working-group sessions with reserve managers from eight central banks. Its central finding, as TFTC summarised it, is that institutions are repatriating physical gold because an asset held in someone else’s vault is no longer considered fully theirs.

The behaviour is already on the record. De Nederlandsche Bank moved roughly 86 tonnes out of New York and Ottawa in September, citing increasing geopolitical unrest: about 59 tonnes sold in New York and replaced with gold bought in London, and more than 27 tonnes physically shipped to its own vault in Zeist. The Banque de France replaced 129 tonnes held in New York with gold stored in Paris. Germany’s Bundesbank still keeps 1,236 tonnes at the New York Fed and says publicly that it has no doubts about the Fed, a position that is becoming lonely.

Gold moved home, and gold still abroad: DNB, Banque de France, Bundesbank

The motive is sanctions risk: Russia’s 2022 freeze made the point that gold inside Russia kept working and gold abroad stopped. That is not the same belief as self-custody, but it is the same logic, arrived at by people who would never describe themselves as cypherpunks: if someone else can say no, the asset is not fully yours.

The same survey work shows how far that logic has to travel before it reaches bitcoin. Of the central banks in the OMFIF Global Public Investor 2026 survey, 92 percent neither invest in nor consider digital assets. Seven percent are considering them, unchanged from 2025. Thirteen percent expect to increase digital asset allocations over the next decade, up from 10 percent the year before. The Czech National Bank remains the only confirmed central bank bitcoin investor, with a $1 million test portfolio bought in November 2025 at Governor Aleš Michl’s direction.

Central bank attitudes to digital assets, OMFIF Global Public Investor 2026

One million dollars against a central bank balance sheet is a rounding error. It is also a data point no other reserve institution has matched, and the gap between accepting the argument and putting it on the balance sheet is exactly the gap this week was about. The argument is won. The allocation is not.

The fix is boring, and it is the only one that works

Insurance is not a fix. A bigger bounty is not a fix. The only durable answer to “somebody else held the key” is to reduce the number of keys that have to be held by anybody, and that work is unglamorous and mostly invisible.

On Saturday a draft specification appeared on Delving Bitcoin for the Multisig Session Protocol. The problem it addresses is old and easy to state: BIP-129 defines the file that describes a multisig wallet, but nothing defines how remote cosigners exchange it, so in practice they use email, chat, or a vendor’s server. In that arrangement the coordinator is the only source of the other cosigners’ public keys, which is the key-substitution problem BitBox described in 2020.

The draft removes the coordinator rather than securing it. Two cosigners share a short code out of band, use it to authenticate ephemeral keys over a relay treated as hostile, commit to nonces in their hellos, then compare a twelve-character string derived from both keys and both nonces by voice or over a verified channel. Mismatch aborts with nothing disclosed. Each side assembles the descriptor itself with a byte-exact function and compares the hash, and each reads the first address off its own signing device before funding. No one distributes a descriptor, so there is no coordinator to compromise.

The author’s own numbers are the interesting part, and he publishes his defects rather than hiding them. The October draft hashed session keys with nothing committed, which let an attacker on the path who already held the code search for a collision at about 2^30 work per side. The November draft commits to a fresh nonce and claims a bound near 2^-59 per session with no offline work. He also concedes that the join is not a PAKE, that a PBKDF2-stretched code is an offline oracle, and that a generated ten-character code carries about 49.5 bits while a typed one does not.

That is what real security work looks like from the inside: a threshold that moved from breakable to comfortably out of reach, a known weakness left visible in the text, two clients supported in version 0.1, and an external design review closed on 18 November. Nothing to post about. Everything that matters.

The same week: Bitcoin Core 31.3 shipped on 20 November. A Delving Bitcoin thread asked whether BIP 54’s new rule at difficulty-adjustment boundaries — a block timestamp must not be more than 7,200 seconds behind its predecessor — creates a slow-sliding median-time-past attack. The author’s own answer is that it requires a majority of hashrate to sustain and would affect one block in 2,016, and he published the question anyway. The protocol’s security is a public argument conducted in the open, one awkward question at a time.

The same week, in another room

Two mathematicians claimed a proof of the Riemann Hypothesis within hours of each other in the last few days, and the shape of that story is the shape of this one.

Levent Alpöge, working with Claude, released Lean code, about 1.3 gigabytes of it and roughly 250,000 pages of equivalent text, rather than a paper. Elliot Glazer confirmed that the formalised statement is the Riemann Hypothesis and put the chance of a Lean bug well below one percent. Within days, Alessandro Masullo released a 940-page human proof he had been working on for twelve years.

Latent Space put the contrast precisely: the machine-written solution is probably correct but incomprehensible; the human one is probably incorrect but understandable.

Sit with that, because it is the same trade this article has been describing. Mathematics has found a way to verify a result without anyone understanding it. The burden moved from authority to proof, and the proof is checkable by a compiler that does not care who wrote it. Bitcoin made that move seventeen years ago with money: you do not have to trust the issuer, the auditor, or the bank, because the rules are checked by everyone who runs the software.

The difference is what happens when the checker is compromised. A Lean proof rests on a compiler and a model that Alpöge did not write and cannot audit. A bitcoin transaction rests on code that thousands of strangers have read. That is not a small difference. It is the difference between verification you perform and verification you delegate — and delegating it is how you end up writing a letter to a thief with a deadline on it.

The honest part

What would make me wrong.

First, Chainflip’s post-mortem. If it turns out that a genuinely distributed threshold of independent validators had to be compromised simultaneously, then my “one weak key” framing is wrong and the situation is worse, not better: it would mean the threshold architecture itself failed rather than being bypassed through onboarding.

Second, the gold story. Eight central banks in confidential sessions, unnamed; the 7 percent and 13 percent figures come from a separate survey of up to 90 institutions, not from those sessions, and I have not seen the primary document. The repatriation is also motivated by sanctions risk, which is a narrower thing than a philosophy of self-custody. I am reading a convergence of logic, not of belief.

Third, BTCS. The $1,450 figure is CoinDesk’s model on the assumption that the collateral balance is unchanged. The company may have added collateral; it may also have withdrawn some to fund the debt reduction, which is the reason the disclosure exists. I do not know which, and neither does anyone outside the company until the next filing.

Fourth, and most uncomfortable: bitcoin at $55,000 did not protect anybody from a 56 percent drawdown. Self-custody protects ownership, not price. If your thesis was that holding your own keys also protects you from volatility, this cycle has been an expensive correction to that thesis.

What I am watching

Monday, 18:38 UTC: whether the bounty deadline passes with a transaction or a statement.

Monday’s US open: whether the liquidation cluster that Glassnode says has been cleared stays cleared, or whether open interest in coins rebuilds past 100,000 with funding still positive. If it does, nothing was purged and the next leg down will find the same crowd in the same posture.

$1,450 on ether: the level at which the code, not a person, decides what happens to the largest listed ether treasury.

The 2027 OMFIF survey: whether the “considering digital assets” cohort moves off 7 percent while gold keeps flying home. If the cohort stays flat while repatriation continues, then the constraint on central bank bitcoin is structural and permanent, not philosophical and temporary.

And one slower thing: whether the coordinator disappears from multisig setup. That is a testable question with a boring answer: whether two people can set up a wallet together without trusting a server, a chat app, or each other’s key distribution. If the answer becomes yes, the surface that produced this week’s losses gets smaller by one seam.

The price will do what it does. The seam is what we can actually close.

Sources

Write a comment