Six Billion Dollars, No Insurance Policy: A Conversation About Proof of Reserves

A proof of reserves is a photograph of a promise, taken by the person who made the promise. After $6.3 billion left Binance hot wallets, the report due Monday can tell you your balance was counted — not that the counting was complete.
the sentence that gets misread

A reserves report is a snapshot, not a solvency opinion

A proof of reserves is a photograph of a promise, taken by the person who made the promise. It can tell you that your balance was counted. It cannot tell you that the counting was complete.

and what it leaves outThe left-hand side of the balance sheet is asserted, not proven.

I have a rule about exchange failures: do not move anything in the first hour. Not because patience is wise, but because hour one is almost entirely noise, and hour two is when an exchange’s real balance sheet starts leaking out through the language of a blog post.

On Thursday night I broke the rule in the mildest possible way. I opened the app, looked at one number, closed the app, and did nothing for forty minutes. That turned out to be the most useful forty minutes of my week, because of what I could not find out.

Here is what is not in dispute. At about 20:05 UTC on 29 October, funds began leaving Binance-controlled hot wallets. TRM Labs counted roughly $6.36 billion across ten networks, most of it inside a twenty-minute burst. Chainalysis put the figure near $6.4 billion. Binance’s own estimate is about $6.3 billion, and it has said it will absorb the loss through its Secure Asset Fund for Users and corporate funds. The FBI has attributed the intrusion to North Korea, and both TRM and Chainalysis name the TraderTraitor group.

The mechanism is the part I care about. Binance’s preliminary account points at a compromised DevOps engineer’s laptop, which led to a service container and to unauthorised withdrawal submissions. Co-CEO Richard Teng was careful to say that “the attacker did not simply use one engineer’s permissions.” Hypernative’s reading is sharper: the signing access was compromised and the signatures were valid, and on Ethereum the attacker used EIP-7702 — the Pectra feature that lets an ordinary account execute contract code — to run attacker-controlled code from a Binance-controlled wallet. Chainalysis describes an operational wallet, funded through Tornado Cash, submitting signed authorisations that let a Binance hot wallet execute code the attacker had deployed.

Read that again and notice what is absent. No private key was leaked, or at least Binance says it found no evidence of one. What moved six billion dollars was not a stolen key. It was a valid signature produced by a system that had been taught to trust its own approval path. If that sentence feels familiar, it should. It is the same sentence as every permit-phishing story, every reentrancy story, and every “the contract did exactly what it was told” story. The difference is only the size of the balance sheet standing behind the signature.

Six Billion Dollars, No Insurance Policy: A Conversation About Proof of Reserves
Aldo Moisio, Public domain, via Wikimedia Commons

I spent the next day in a long argument with a friend — call him R — who keeps a working balance on the exchange because he trades, and who wanted a yes or no. What follows is that argument, compressed and cleaned up.

R: They said users are safe. So I’m safe. Yes or no.

Me: Both. You are safe from the loss and exposed to the mechanism, and those are different sentences that people keep merging into one.

R: What does that even mean.

Me: When the coins left at 20:05, your balance did not change. It is still a row in a database that says the exchange owes you something. What changed is the thing standing behind the row. Before Thursday it was a pile of coins in a wallet. After Thursday it was a promise plus a shopping list.

By Friday, Binance said it had reserved more than $6.34 billion in company funds to cover the gap, of which roughly $5 billion was already held in the same tokens that were stolen — USDT, USDC, bitcoin, ether, BNB — covering about 80% of the replacement, with the remainder being bought on the market and over the counter. That detail is the whole custodial business model in one line, and almost nobody reads it that way. The exchange was already long the assets it owed you. The “rescue” was not new money arriving from outside; it was inventory being reassigned inside the same building.

R: Fine. But the proof of reserves will show my balance is there. That’s the point of it.

Me: That is exactly half the point of it, and the half that was never the problem.

sum(balances the exchange says it owes) <= sum(coins it controls) at one block height

the only inequality a reserves report can assert

That inequality is what a reserves report is trying to demonstrate. The machinery behind it is usually some form of Merkle tree of user balances: every customer’s leaf is hashed into a root, the exchange signs the root, and each customer can check that their leaf is inside the tree without seeing anyone else’s. Binance has published this kind of report since 2022, and the cryptography is not the weak part. The weak part is the left-hand side of the inequality.

R: The left side is the liabilities. They know their liabilities. They’re an exchange.

Me: They know their customer ledger. That is not the same object. A reserves report proves that the leaves the exchange chose to include are inside the tree. It does not prove that every liability is a leaf. Negative balances, margin loans, internal settlement accounts, sub-accounts that roll up somewhere else, the difference between what a user sees and what the ledger says after netting — all of that is on the left-hand side, and none of it is in the tree unless the exchange decides to put it there. This is not an accusation. It is a definition. A proof of reserves is a proof of assets plus an assertion of liabilities.

What the report actually shows

A signed root at a specific block height, and a leaf for your account. Your balance was counted, at that moment, by that institution, with that tree.

What it cannot show

That the tree is complete. That the coins are unencumbered. That the block height was not chosen. That no one at the exchange can produce a second, friendlier tree for a different audience.

What would actually settle it

A liabilities-proof where each customer signs their own balance into the total, so the exchange cannot quietly omit anyone; plus an audit opinion with a firm’s name and liability on it; plus the boring one, which is withdrawals working.

The third item on that list is doing more work than the first two, and it is the only one you can verify yourself. Which brings us to the number everyone quoted and nobody priced.

Freeze Is Not Symmetric

By Friday, roughly $3 billion had been frozen on-chain, mostly stablecoins locked by their issuers, with Elliptic independently verifying about $2.84 billion. That sounds like the story is two-thirds solved. It is not, and the reason is the most useful thing a customer can learn from this week.

TRM’s snapshot, taken at 00:30 UTC on Friday, still had about $2.4 billion with the attacker in ether and bitcoin — more than 420,000 ETH and 18,270 BTC — which no issuer on earth could freeze. The attacker was not careless. Chainalysis traced about $500 million of stablecoins being moved into a wallet for swapping, and over roughly twenty minutes $325 million of stablecoins became $285 million of ether. That twenty-minute window is the whole security model of every stablecoin issuer, described from the outside: the freeze only works while the money is still the kind of money that has an admin key.

BucketRough sizeWho can actually stop it
Stablecoinsabout $3 billion frozenThe issuer, through a blacklist no holder agreed to
Ether and bitcoinabout $2.4 billion still movingNobody. Not the exchange, not the FBI, not the issuer
Company funds committedabout $6.34 billionBinance’s treasury, which is not audited like a bank’s

R: So my USDT is safer than my bitcoin.

Me: Your USDT is more reversible than your bitcoin. Those are not the same word, and the difference is who holds the switch. If you hold USDT on an exchange and the issuer freezes the exchange’s address, your balance may keep showing the right number while the thing behind it has been frozen by a company you never signed a contract with. That is a feature when the money is being stolen. It is the same feature when the money is yours.

If you want the longer version of that argument — whether a public chain should honour a freeze request at all — I wrote it up around the Bitget theft, which now turns out to be connected: Chainalysis says funds from that $387 million incident commingled with the Binance funds, and ZachXBT says the same actor is likely behind both. The uncomfortable conclusion has not changed. Reversibility is a property of the asset, not a virtue of the exchange.

Six Billion Dollars, No Insurance Policy: A Conversation About Proof of Reserves
Anonymous, Public domain, via Wikimedia Commons

Who Actually Pays

Four numbers circulated this week. They are worth separating from the sentences attached to them, because each one is a claim wearing a measurement’s clothes.

NumberWhat it isWhat it is not
$6.3 billionBinance’s estimate of what leftA settled loss figure. Chainalysis says $6.4 billion; TRM counted $6.36 billion across ten networks
$3.0 billionFrozen on-chain, mostly stablecoinsMoney returned to users. It is evidence sitting in someone else’s wallet, pending legal process
$2.4 billionETH and BTC still with the attackerRecoverable by anyone’s order. There is no admin key for bitcoin
$6.34 billionCompany funds committed to cover the gapAn audited capital ratio. It is a commitment, made on a blog, in a week
$6.3B
estimated outflow, the largest crypto theft on record

R: Where does the money come from, then.

Me: Equity. Reuters reported, citing people familiar, that Binance carries no insurance for this, so the loss lands on corporate reserves and SAFU rather than on a policy. That is exactly what bank capital is for, and it is worth being precise about the difference. A bank holds capital against risk-weighted assets, reports a ratio to a regulator every quarter, and can be resolved by an authority that has a legal right to wipe out shareholders and haircut depositors. An exchange holds reserves, publishes a proof of reserves, and can be resolved by a blog post that says it will absorb the loss. Both are promises. Only one of them has an institution standing behind the promise whose job is to break it when it fails.

R: You’re describing a bank. Binance isn’t a bank.

Me: Correct, and that is the entire point. It is doing a bank’s job with none of a bank’s plumbing — no deposit insurance, no capital ratio, no resolution regime, no supervisor with subpoena power. What it has instead is a fund, a treasury, and a very large Twitter account. None of those are bad things. All of them are unverifiable from outside, which is the same complaint I have about the reserves report.

The Part Nobody Says Out Loud

The weirdest sentence of the week was not about the hack. It was the co-CEO noting that the $3 billion frozen represents about 2% of total user assets and public reserves. On one reading that is reassurance: the hole is small relative to the balance sheet. On another reading it is an admission that the number you should be watching is not the loss at all — it is the denominator, and you have no way to check it.

That is the honest summary of custodial crypto in 2026. The chain gives you perfect visibility into every coin that moves, and the institution gives you a signed root that tells you almost nothing about whether the coins it claims to hold are the coins it actually holds. The visibility is real. It just points at the wrong layer.

Six Billion Dollars, No Insurance Policy: A Conversation About Proof of Reserves
bradhoc, CC BY 2.0, via Wikimedia Commons

What I Actually Do Now

I am not going to tell you to self-custody everything, because for most people that trades a custodial risk for an operational one and calls it safety. Here is the smaller, duller version that I have actually changed.

  • Split the question in two. “Is my balance on their books” and “can I get my coins out today” are different questions, and only the second one is a solvency test. A reserves report answers the first. A completed withdrawal answers the second, and it is the only answer that costs the exchange anything.
  • Exercise the exit path on a schedule. One small withdrawal a month, and note how long it takes and whether anything gets held for review. The value is not the coins; it is that you find out what the queue looks like while the queue is short. Everyone discovers their withdrawal limits during the one week when limits matter.
  • Price the asymmetry, don’t moralise about it. Stablecoins are reversible by an issuer you have no contract with; bitcoin is reversible by nobody, including you. That has consequences for where each one should sit — and “on an exchange, because it’s convenient” is a choice about which failure you are willing to underwrite.
  • Treat every approval as the object of the attack. The six billion did not leave through a broken cipher or a leaked key. It left through a legitimate approval path that had been made to approve. If you want the mechanics of why a valid signature is not a valid consent, that is the piece I wrote on EIP-712, and this week was that piece with a bigger balance sheet.

The reserves snapshot was due Monday. Look at it, by all means. It will be signed, it will be large, and it will be true at one block height. The only thing a proof of reserves has never been able to prove is the thing you actually want to know: whether the person counting is counting everything.

(The End)

Blockchain

One Withdraw, Paid 29 Times: Reentrancy, Read Line by Line

2026-10-30 2:35:43

Blockchain

Nine of Ten Agents Raised Their Own Spending Cap. The Approval Was Real. The Human Wasn't.

2026-11-2 3:14:07

❯
Profile
Cart
Coupons
Search