Decent Newsroom
Sign In

Nobody Froze the Bitcoin

By Roger October 31, 2026
A $6.3 billion theft, a freeze that reached only the assets with an issuer, and a protocol that answered AI-assisted bug hunting with a shorter embargo instead of a rescue.
Nobody Froze the Bitcoin

Nobody Froze the Bitcoin

At 20:20:06 UTC on Thursday, block 973,732 was mined with 3,898 transactions inside it. Buried in the middle were two chains of transfers that step down by exactly 1,000 BTC per hop - 7,636.80, 6,636.80, 5,636.80, 4,636.80, 3,636.80, 2,636.80 in one, and 4,696.60, 3,696.60, 2,696.60 in the other. Somebody was moving coin in round thousand-coin slices, ten times in a row, inside a single block. And in that same block, one address - bc1q0mtc2f5agpkvlq6hhahmfr3hlr55kyfjhl2mnm - received eleven outputs that add up to exactly 18,270.20 BTC.

I checked that on my own node rather than reading it in a headline, because 18,270 is a number worth verifying before you build anything on top of it. It is there, to the satoshi. Someone who had every key they needed signed a ladder of transactions that emptied a wallet into a fresh address in about ten minutes, and the network did exactly what it promises: it accepted them.

The rest of the story happened off-chain, and that is where the real news is. Binance lost roughly $6.36 billion across ten blockchains in what TRM Labs calls the largest crypto theft on record, most of it inside about twenty minutes of the first transfer at 20:05 UTC, and about four times the $1.5 billion that left Bybit in February 2025. Within hours, stablecoin issuers froze somewhere between $2.7 billion and $3 billion of it. Nobody froze the 18,270 BTC. Nobody could.

That asymmetry is the most important thing that happened this week, and it is not a story about North Korea, or about artificial intelligence, or even really about Binance.

Block 973,732 - the 1,000 BTC peel chains and the eleven outputs to one address

What signed

Binance’s own first statement is the part to read slowly. Its initial investigation, in the company’s words, “suggests that the security incident may involve a compromised DevOps engineer’s laptop, which allowed an attacker to gain access to a service container and submit unauthorized withdrawal transactions. We have found no evidence of any wallet private key leak.”

Six hours later, co-CEO Richard Teng went further in a video update. This was not one person, he said; the investigation had found additional points of breach; the attackers ran a self-destruct mechanism that destroyed their tooling and traces as they went. And on Ethereum, they used EIP-7702 to take over a hot wallet, where “the delegation did not pass through our normal approval controls.”

EIP-7702 arrived with the Pectra upgrade in May 2025 and lets an ordinary Ethereum account hand control of its behaviour to a contract by signing an authorisation. TRM’s reconstruction is that the attacker prepared fifty such authorisations in advance and submitted them in a single transaction at 20:20 UTC, converting one of Binance’s best-known Ethereum hot wallets into a sweeper only the attacker could call.

Sit with the shape of that. These wallets lived on ten chains with ten different key types and signing schemes, and every one of them produced a valid signature for the same person within half an hour. TRM’s best explanation is not that one private key leaked. It is that the attacker gained the ability to make Binance’s hot wallet signing system sign whatever it wanted. Not one vault, but the machine that tells every vault to open.

Marty Bent put the plain version in his newsletter the same day: “You should not have $6 billion of other people’s money sitting behind a setup that one compromised laptop can talk to.” He also noted the seven-year echo - Binance lost 7,000 BTC from a hot wallet in May 2019 to phishing and malware, and covered that loss from its own balance sheet too. The soft spot never moved. It just got bigger.

The cold storage is the tell. The address identified as Binance’s cold storage held roughly 248,600 BTC on Friday and has not moved. Whatever was reachable by the signing service went; whatever sat behind a slower, more annoying, more manual process stayed. That is not luck. That is architecture, and it cost somebody a few minutes of inconvenience per day for years, which is exactly the kind of cost that gets optimised away.

The market’s verdict

Bitcoin closed at $81,267.59 on 28 October. It closed at $77,775.11 the next day, the day of the theft, and $77,198.21 the day after - a 5.01% drawdown from the pre-hack close, measured on daily closes. On Saturday morning it was trading around $77,240, down 4.96% from where it started the week.

Bitcoin daily closes, 1-31 October 2026

Anthony Pompliano’s read is that this is the story: “a decade ago the digital currency would have plummeted 20-40% on this type of news.” I cannot verify the decade-ago counterfactual, and I will not pretend otherwise. What I can verify is the current number, and five percent for the largest theft in the industry’s history is a very small number.

Two things are doing the work there, and only one of them is confidence. The first is that the loss is not bitcoin’s loss. Eighteen thousand coins moved to an address that is now publicly watched; the supply did not change, no rule was bent, no issuance was touched. Traders who understand the difference between a protocol and a company priced it correctly.

The second reason is duller and more important: Binance said it had set aside more than $6.34 billion to cover customer losses, and its shareholders have covered losses before. A theft the exchange eats is a transfer from a balance sheet to a user base. It is bad for the equity and roughly neutral for the asset. That is a very different event from 2022, when the question was whether the customer would be paid at all.

The switch

Now the part that most coverage skipped past. Between $2.7 billion and $3 billion of the haul was frozen within hours. Tether and Circle and others moved faster than any court could have, and the reporting is honest that this is the only reason a large share of the money might come back. Two figures are circulating - roughly $2.7 billion from TRM-derived reporting and about $3 billion in Binance’s own statement - and I will give both rather than pick the tidier one.

Consider what a freeze is. It is an administrative act, taken by a private company, on a Saturday, against funds in wallets the company does not control, executed by refusing to honour a redemption. It worked because a stablecoin is not a bearer asset. It is an IOU with a phone number attached. Whoever answers that phone can decide, without a hearing and without a statute, that your token is no longer yours to spend.

For the victims of this hack, that power is the difference between a loss and a recovery, and I am not going to pretend otherwise. But notice who holds it. CoinDesk’s reporting on the freezes includes the question that matters more than the mechanics: legal experts asking whether a smaller victim - a person, a small exchange, a startup - would get the same protection. The answer, obviously, is no. There is no procedure to petition. There is a compliance department and a press release.

And notice what could not be frozen. Twenty-two point six percent of the dollar loss, by my arithmetic from CoinDesk’s $1.4 billion valuation and the $6.36 billion total, was bitcoin, and it is the only portion where the thief currently holds the asset outright with no intermediary in the path. Every dollar the industry recovered this week was recovered from an asset that had an issuer. Every dollar it could not recover was in an asset that does not.

That is not a bug in bitcoin. It is the entire design, seen from an uncomfortable angle. The same property that let a stranger move 18,270 coins against the wishes of the largest exchange on earth is the property that makes the coins worth holding. Settlement finality does not ask whether you are a thief. It asks whether you signed.

The steel-man

The strongest case against my framing deserves its own paragraph, because it is not weak. If you are a Binance customer on Thursday night, the thing you want is not immutability. It is a phone number that answers. The freeze worked. The exchange backstopped the rest. Nobody’s rent money disappeared, and the reason is precisely the concentrated, permissioned, discretionary machinery that cypherpunks spent thirty years trying to avoid. A world of pure bearer assets would have delivered a total loss with beautiful cryptographic elegance.

There is a second version of the argument that is harder to dismiss. Bitcoin’s immutability is not a neutral property; it is a property that protects whoever currently holds the keys, and after a theft the current holder is the thief. The same ledger that makes confiscation impossible makes restitution impossible. Analysts can label an address, but labels do not move coins. The only remedies available - sanctions, exchange blacklists, stablecoin freezes - are all interventions in the fiat and stablecoin layers, not in bitcoin. The asset’s purity is what makes it valuable and what makes it useless when you need it to be a court.

I hold both of those to be true. What I reject is the conclusion people draw from them, which is that the trade-off can be dodged by picking better custodians. Binance is not a badly run exchange in the way a startup with two engineers is badly run. It is the largest one, with a security budget to match, and it still had a signing service one laptop could reach across ten chains. The failure mode is not incompetence at the margin. It is the natural end state of any system where convenience is priced at zero and delay is priced as a cost.

What I could not verify

Four things stayed open, and I would rather name them than smooth them over. Pompliano’s comparison to a 20-40% drawdown a decade ago is a claim about a counterfactual I have not tested; my five percent is measured, his alternative is not. The size of the freeze is reported two ways, roughly $2.7 billion and about $3 billion, and I could not reconcile the gap from public sources, so both figures stand in the text. Nobody has shown that an AI model played a role in the Binance breach, and the strongest reporting attributes it to a state-linked group working through compromised infrastructure — the AI connection in this piece is about the industry’s second-order response, not about the entry point. And Binance’s account of how the signing system was reached is, so far, Binance’s account; the company says its own investigation found additional points of breach, which means the version published on Thursday night is already superseded.

Black October

Then there is the second half of the week, which almost nobody has connected to the first.

Nathan Lambert coined the term “Black October” for the run of critical vulnerability disclosures - PHP2Shell, GitLab - that marked the moment the cyber risk of AI models started to arrive in the real world. His argument, and he is candid that he may be wrong, is that the open-weight doomsaying was a falsifiable prediction and it is being falsified politely: GLM-5.3 is the model that crosses the capability threshold the safety community described, its weights are out, and a month later the world looks like an acceleration rather than a step change. Epoch AI used GLM-5.3 to build a website that, when a Firefox user merely visited it, handed the attacker file exfiltration. Real, serious, and one step in a multi-stage attack that still needs a delivery channel.

The most concrete institutional response to that came not from a government but from a mailing list. Bitcoin Core amended its security policy for vulnerabilities it classifies as easily discoverable, “such as those found with a simple prompt to a large language model.” The project had seen a rising number of such reports, which can be rediscovered or inferred from covert fixes. Instead of waiting for the next major release and the severity-based embargo, it now backports the fix to every maintained release, cuts point releases as soon as possible and publishes an advisory with them. The three vulnerabilities in this week’s notable code section were disclosed under the new rule. Their fixes merged after 32.0 had been tagged, so 32.0 was never released at all; 32.1 shipped instead, along with 31.2, 30.4 and 29.5.

Read that again as a piece of institutional design. An AI model made bug discovery cheap, so the disclosure pipeline got shorter and the release cadence got faster. No emergency funding round, no regulator, no freeze. A policy change and a patch. Compare it with the CLN denial-of-service bug disclosed this week - found with a fuzz target, reported in April, fixed in 26.04.1, written up six months later, with the caveat that upgrading alone did not remove the poisoned gossip message, an operator had to delete it by hand. Unglamorous, boring, working.

And the central bank has already put a number on the ambient risk. Philip Lane’s speech at the San Francisco Fed on Friday notes that IT, cybersecurity and related incidents account for 10-12% of all operational risk events at banks in 2025, nearly double their incidence in 2018. That is ECB survey data, not a vendor’s scare chart.

The other side of the ledger

Lane’s speech is also the most useful document of the week for anyone trying to size the AI economy, because it separates the capex from the adoption. Euro area digital investment has grown 73% in real terms since 2014, from 2.3% to 3.3% of GDP. The United States went from 3.3% to 7.3% in the first half of 2026 - more than double, and more than twice Europe’s level. In the ECB’s corporate telephone survey, firms report that AI has added about 10% to their capital expenditure plans this year.

Then the adoption numbers, which are much softer than the capex numbers. In the fourth quarter of 2025, 64% of firms in the Survey on the Access to Finance of Enterprises said they used AI experimentally or moderately, and only 7% described their use as significant. Among workers, 52% report using AI at work, up from 26% in 2024, and the median user saves three hours a week - 7.7% of median working time. Only 15% cite reducing labour costs as a reason to use it; 62% cite improving business processes.

The capex gap and the adoption gap - ECB survey data

There is one finding in there that should interest anyone who watches markets for a living. Firms plan more AI investment when they expect more domestic peers to invest, and the effect is large: a one percentage point increase in the expected share of domestic firms investing raises a firm’s own expected investment rate by about 0.6 percentage points, against an average expected rate of 8.3%. Beliefs about foreign firms’ plans have no effect at all. That is a herding coefficient with a number attached, and it is measuring exactly the mechanism that turns a capex cycle into a capex bubble.

Which brings us to Friday’s other headline. Google’s chief business officer told investors that “the strong growth in AI Mode queries had an impact because of the lower monetization we were seeing in AI Mode” - the company’s own AI product is cannibalising the revenue per query of the search product it replaces. Alphabet’s shares fell about 15%, roughly $500 billion of market value, the worst reaction since its 2004 listing according to Latent Space’s summary, and the Nasdaq 100 is down about 10% in three weeks from its 6 October record, including 4.5% on Friday alone. OpenAI is reportedly raising at a $1.25 trillion valuation rather than $1.4 trillion.

And the buildout is running into the ground it stands on. Finnish police opened a criminal investigation on 29 October into Google’s subsidiary Tuike Finland Oy over suspected environmental law violations at its Muhos data centre site. According to the Oulu Police Department, Tuike cleared roughly 330 hectares near Muhos and about 200 hectares near Kajaani before the required environmental impact assessments were completed; the supervisory agency had already ordered work halted at both sites on 6 October. Google committed $15 billion to Finnish AI infrastructure in September and described it as its largest investment in Europe. The legal theory now driving the probe is the Water Act, tied to the drainage work - roughly nine kilometres of new ditches, most of them dug before the assessments were done.

Put the two halves of the week next to each other and the same sentence appears in both. For fifteen years the industry’s core assumption was that the abstraction in the middle is free: that you can hold a bearer asset through a custodian without paying for the custodian’s failure mode, and that you can wrap a business model in a model without paying for the wrap. This week sent invoices for both. Binance’s was $6.3 billion and a suspended withdrawal service. Google’s was a monetisation hole it had to describe out loud, in a sentence composed by its own chief business officer.

The ECB’s forecasters, for what it is worth, are not pricing a collapse. The Survey of Professional Forecasters published Friday has euro area inflation at 3.0% in 2026, 2.6% in 2027 and 2.0% in 2028, with longer-term expectations anchored at exactly 2.0%, and growth at 1.0%, 1.3% and 1.3%. Wage growth of 3.3% this year. Nothing in that table looks like a regime break. It looks like slow, sticky, unexciting inflation that a central bank can live with, which is its own quiet argument against the assets that exist to hedge the alternative.

What the protocol did instead

Here is the contrast I keep coming back to, and it is 371 bytes long.

On 29 October, the same day the ladder left Binance, someone posted to Delving Bitcoin a multiplication routine for Tapscript that fits in 371 bytes and requires no soft fork. OP_MUL has been disabled since 2010. Under BIP-342, any product has to be assembled from additions and shifts, and every intermediate that a later opcode reads must still fit in four bytes. The tightest symmetric bound that survives that rule is 46,340 - because 46,340 squared is 2,147,395,600, just 88,047 under the limit, and one more fails.

The construction is witness-assisted: the spender supplies the sixteen bits of the multiplier, the script walks them from most to least significant while maintaining the running product and a reconstructed multiplier, and a five-byte epilogue verifies that the reconstruction matches what was parked on the altstack. Fifteen 23-byte rounds plus a shorter final round plus a prologue and an epilogue. Thirteen active opcodes. A 793-line Rocq proof with six closed theorems and no admitted lemmas, plus a Rust auditor and a Python reference, because the author found two real bugs only when testing against the actual interpreter - an OP_MAX where OP_GREATERTHANOREQUAL was needed, and 46,341 encoded as 46,325.

Nobody voted on it. Nobody froze anything. Nobody had to be rescued. A capability the protocol deliberately removed in 2010 was rebuilt inside the existing rules, with a machine-checked proof, and deployed by whoever wanted it.

That is the thing the headlines miss when they call this a security story. The part of this industry that answered the week’s hardest question - how do you add power to a system without adding a person who can be compromised - answered it with arithmetic. The part that answered it with a phone number lost six billion dollars, froze what it could, and promised to cover the rest from the balance sheet.

What to watch

Three things, in order of how much they would tell us.

Whether Binance publishes its signing architecture. Teng promised a full review across “endpoints, cloud, key management, signing, everything,” and the interesting part is not the audit but the design that replaces the sharing. If the fix is more monitoring on the same single signing path, the next laptop wins too. If it is per-chain isolation with human quorums and hard delays, the cost will show up as slower withdrawals and customers will complain about it, which is the only honest signal that a real control was installed.

Whether the freeze becomes a policy. Right now the emergency power of the stablecoin issuer is exercised case by case, by firms whose own discretion is the whole procedure. The victims of this hack got a rescue. The next smaller victim will get a support ticket. Watch for any issuer publishing criteria - not promises, criteria - for when it will freeze.

And whether the market keeps pricing custody as a footnote. Five percent for the largest theft in history is a confident market. Confidence is cheap while the balance sheet covers the loss. The question is what happens the first time it does not, because the entire argument for a $6.34 billion backstop is that Binance is profitable enough to have one, and that is a balance-sheet argument, not an architectural one.

The ladder in block 973,732 is still sitting there, 18,270.20 coins deep, watched by everyone and reachable by no one. It is the most honest monument this industry has built: proof that the asset works exactly as specified, and that almost nobody wants to live with what that means.

Sources

Write a comment