Decent Newsroom
Sign In

Yes or No Is Not an Answer

By Roger November 1, 2026
Six billion dollars moved through chains that can freeze, twelve DeFi sites fell with one registrar, one passcode cost a man fourteen thousand dollars — and the only thing that got more capable all week was a ten-byte script on a test network.
Yes or No Is Not an Answer

Yes or No Is Not an Answer

A developer spent this week rebuilding the same small payment channel for the third time, and the sentence he used to explain the difference is the most useful thing I read all week. Under the old design the script asked the network a question and got a verdict. Under the new one, he wrote, “the script gets the hash back instead of just a yes or no.”

That is a small change inside a test network that most people will never touch. It is also the exact distinction that separates the five things that went wrong in the same seven days.

Because look at what we were told this week, and notice that every single statement is a yes or no. Ninety percent replenished. One hundred percent coverage. No indication of sensitive data. Twelve domains, one registrar. One passcode. Five thousand dollars for an account takeover.

None of those are answers. They are verdicts, and a verdict you cannot check is a rumour with better grammar.

Ninety percent of what

Binance says it has replaced 90% of the cryptocurrency stolen on Thursday, up from about 80% in its previous update. Customer holdings of bitcoin, ether, USDT, BNB, SOL, USDC, XRP, USD1, RLUSD and ZEC are, in its words, “now fully backed.” The exchange has set aside more than $6.34 billion in company funds, including $5 billion held in the affected cryptocurrencies, and plans to restore its Secure Asset Fund for Users to roughly $1 billion. Withdrawals were due to reopen on Sunday.

Read the last paragraph of that reporting instead of the first. Binance published a proof-of-reserves snapshot taken at midnight UTC on 31 October showing coverage ratios at 100% or more for the named tokens. The full cryptographic proof — the one that would let a user verify that their own balance is inside the tree of liabilities Binance has committed to — is expected by 3 November. It does not exist yet.

That gap is not pedantry, and Binance is not being unusually coy. It is the normal shape of the genre. A reserve snapshot answers “do the addresses we showed you hold enough?” It does not answer “are those the only liabilities?” and it does not answer “do we control the keys to those addresses?” A proof of liabilities with inclusion proofs, published on a schedule, signed by an auditor whose method you can read, answers more. Almost nobody publishes that, because it is expensive and because it makes the balance sheet permanently auditable by strangers.

And in this specific case there is a harder question sitting underneath. CoinDesk’s review of Binance’s earlier reserve disclosures found that wallets drained in the attack were among those the exchange had listed as backing users’ holdings. The attestation was correct in form. It was wrong in substance within minutes, and no user could have known until the theft made it visible.

No indication

On Friday evening, Chainalysis published a notice to its own customers. An unidentified attacker had broken into its systems and internal tools this month, the company said, using “offensive agentic techniques,” and had accessed “some non-sensitive customer data, such as user IDs, full names, email addresses, and organization role labels.”

The follow-up sentence is the one worth keeping: based on the evidence available to date, the firm said, it had found no indication that the incidents resulted in unauthorised access to investigation content, transaction or query results, passwords in clear text, MFA secrets, or user session tokens.

I have no reason to doubt them. I also want to be precise about what that sentence can and cannot carry. “No indication” is what an investigation reports when it has not yet found the thing, and it is the honest phrasing — but it is a statement about the state of their knowledge, not about the state of the world. The distinction matters because the entity making it is a firm whose product is the ability to trace money. If anyone should be able to answer with a value instead of a verdict, it is Chainalysis.

The same notice contains the week’s best line about the laundering problem, from CEO Jonathan Levin. The attackers had already used “thousands of addresses,” he said, and “we anticipate this will become tens of thousands over the coming weeks.” Then the reason, which is almost economic: “They have too much money for their trusted laundering venues to handle.”

That is a liquidity constraint on crime. Six billion dollars is more than the specialist money-laundering market can absorb at the speed it wants to move, so it spills into bridges, instant exchanges and swappers — venues with thinner compliance and worse hygiene, and therefore more places for an analyst to catch it.

On Saturday the FBI put a number on the other side of that. It confirmed that North Korea’s TraderTraitor group, a subunit of Lazarus, was behind the theft, and listed the addresses it wants the industry to block: over 600 Ethereum addresses, just under 20 Bitcoin addresses, and a single Solana address. It asked THORNode operators, RPC node operators, instant exchanges, DeFi services and analytics firms to refuse transactions “with or derived from” those addresses.

Sit with that ratio. More than thirty Ethereum addresses for every Bitcoin address, and one for Solana.

Where the FBI’s address list points - addresses named by chain

That is not a story about which chain was robbed. It is a story about which chain a $6 billion laundering operation can physically use. Bitcoin’s ledger is public, permanent, UTXO-shaped, and served by a comparatively small number of venues with heavy chain-analytic coverage. Ethereum and the stablecoin layer offer programmable money, deep liquidity, and issuers who can freeze. The FBI’s list is short on Bitcoin because at that scale Bitcoin is a bad place to run a laundry. CoinDesk reported that stablecoin issuers had already frozen more than $2.4 billion of the haul, which is a sentence no Bitcoin address can generate.

The allowlist

Now move from money to a chat window, because the mechanism is identical and the details are better documented.

Parameter Labs published an exploit chain this week against ChatGPT’s HTML preview feature — the sandboxed iframe where the model renders interactive apps. Simon Willison summarised it, and the summary is worth reading slowly because it is a textbook lesson in how composition defeats policy.

OpenAI uses a library called MathLive for graph support. MathLive ships its own virtual keyboard, and that keyboard can be triggered from the parent page by a postMessage() call from inside the sandboxed frame. The receiving code on the parent side uses innerHTML. So anything that can send a message can add markup to ChatGPT’s own page — and nothing prevented other sandboxed frames, including HTML preview frames, from sending that message.

innerHTML will not execute a <script> element, which is why this is not immediately fatal. But it will happily insert a new iframe, and an iframe inserted that way does not inherit the sandbox restrictions of the one that requested it.

OpenAI’s Content-Security-Policy blocked inline script entirely and allowed script loading from only a handful of domains. One of them was YouTube. And YouTube’s oEmbed endpoint could be made to reflect JavaScript back in a form that executed.

After that, it was routine: script running in ChatGPT’s origin can fetch the signed-in user’s private data and exfiltrate it through an image reference. The result was a share URL that, when a logged-in user opened it, could read their data or act as them. OpenAI closed the hole and paid a $5,000 bounty. Willison’s verdict — “a bit cheap for a vulnerability this serious” — is, I think, correct.

Every layer here made a defensible decision. Sandbox the iframe: yes. Whitelist a few script origins: yes. Use innerHTML in a keyboard handler written years earlier: careless, but common. The failure lives in the composition, and the composition has no owner. That is what an allowlist really is: a list of the exceptions your product team needed, granted years apart, each individually reasonable, jointly a script loader.

One account, one passcode

The two smallest stories of the week are the same story.

The Block reported that twelve DeFi websites were hijacked and about $138,000 stolen, and that all twelve domains were registered through Name.com, according to registry records the publication accessed — a pattern that points at a compromise upstream of the sites rather than at twelve separate breaches.

Twelve projects, one registrar account. The smart contract may be audited, immutable and formally verified; if the domain resolves to somebody else’s front end, the contract is decoration. Users do not read bytecode. They click the bookmark, connect the wallet, and sign whatever the page asks for. The registrar is a custody layer that nobody in this industry lists as a custody layer, and it is protected by a password and, if you are lucky, a hardware token.

Then there is Magnus Ahmad, who lost about $14,000 in a few hours in Hackney in August, in a case CoinDesk reconstructed this week. Two men spent an evening being friendly in a bar and watched him type his iPhone code. Outside, a third took the unlocked phone and ran. Crypto started leaving his MetaMask wallet at 1:53 a.m. Phantom and Bybit were accessed. By 3:22 a.m. six unauthorised card transactions had cleared, including more than $2,000 at a convenience store. His eToro account was used to open leveraged positions representing nearly $44,000 of exposure. His total loss: about $9,500 in crypto, nearly $3,000 in card charges, and more than $800 in trading costs.

The passcode was never designed to be a master key. It became one, because the phone is where the email is, and the email is where the password resets are, and the wallet apps sit two icons away from the bank.

What actually got smaller

Against all of that, the week’s only unambiguous progress happened where nobody was looking, in a Delving Bitcoin thread with one post and one participant.

The author had built the same Alice/Bob/Carol channel three times. First with ANYPREVOUT plus CTV, then with a CSFS rekey ladder, and now with a third construction made possible because Inquisition 29.4 — the test network where proposed opcodes get switched on before they are anywhere near Bitcoin — enabled OP_TEMPLATEHASH. He used the scripts from a draft BOLT proposal.

The new update script is three opcodes and comes to 10 bytes:

TEMPLATEHASH INTERNALKEY CSFS VERIFY <500000000+n> CLTV

The settle branch is 35 bytes: TEMPLATEHASH <hash> EQUAL. The witness for an update is 143 bytes and stays 143 bytes however many times the channel updates. The CSFS-plus-CTV ladder needed 42 bytes of script and a witness that started at 512 bytes and grew with every rekey. The ANYPREVOUT version was leaner — roughly 35 bytes of script and about 150 of witness — but it needs APO, and the whole point of the exercise is to find paths that do not.

The witness that stopped growing - bytes per channel update, three designs

He then did the thing that makes a demo into an argument: he published an old state on purpose, and let a newer update replace it. State three was signed before any state output existed. That is the property eltoo exists for — any later state can override an earlier one, so a counterparty broadcasting a stale state loses rather than wins, and the punishment machinery disappears.

Now the sentence from the top of this article. Why is the new script three opcodes? Because of what the node hands back. With ANYPREVOUT, the rebinding happens inside signature checking and the script only hears “valid” or “not valid”. With CTV, the script gets a yes or no and never sees the hash, so to get CSFS to sign it you have to carry a second copy of that hash in the witness yourself. With TEMPLATEHASH, the node puts the hash on the stack. CSFS signs whatever is on the stack. Three opcodes.

The byte count is the small part of that. The structural part is that a value on the stack can be composed with other opcodes, and a boolean cannot. One design ends the conversation; the other hands you a number you can build with.

Caveats, because this is a toy: one key, no HTLCs, four transactions including the deliberate stale state. It is a proof of a shape, not a payment channel. But it is the only thing in this entire week that made the system strictly more capable, and it was published for free by someone with nothing to sell.

Two balance sheets, one price

Then the part that should worry the people running treasuries.

Next Technology Holding, a Japan-based software company listed on Nasdaq, told the market after Friday’s close that its board had authorised selling or borrowing against roughly 5,833 bitcoin, worth about $449 million. Its stated reasons: the Binance hack, exchange vulnerabilities, and the potential threat quantum computing poses to bitcoin’s cryptography. The company says none of its bitcoin is held at or through Binance.

Look at the rest of the filing. Next Technology closed Friday at $5.30. With about 1.47 million shares outstanding, that values the entire company at roughly $7.8 million. Its bitcoin, by CoinDesk’s arithmetic, is worth more than fifty times that — fifty-seven times, if you divide 449 by 7.8. And the filings reviewed by CoinDesk do not identify the institutional custodians holding “substantially all” of the coins, or provide wallet addresses that would let an investor check the balance.

So a board is preparing to sell an asset it cannot prove it holds, citing a risk from an exchange it does not use, and a cryptography threat that is real in the literature and not imminent in any measurable way.

I understand the decision better than I understand the explanation. A company whose reported bitcoin is 57 times its equity value is a company where the market has already decided not to believe the number. When the market does not believe you, you do not get to wait for a better price — you de-risk before the doubt becomes a run. The Binance hack gave every such board a reason to act that sounds like diligence and functions as cover.

The second filing is cleaner, and sadder. Iole Inc., listed on the Tokyo Stock Exchange’s Growth Market, resolved on 30 October to sell its entire 151.21 bitcoin by the end of November, expecting ¥1.831 billion in proceeds against a book value of ¥2.085 billion — a realised loss of ¥254 million, about 12.2% of the position’s carrying value. The money goes into AI data centres, AI computing services and a segment the company calls “Neo Cloud.” It is keeping its Hyperliquid tokens, accumulated since July.

Bitcoin goes. The altcoin stays.

That single line tells you which cohort Iole belongs to, and it is not the cohort that reads the white paper. It is the cohort that bought a narrative, and has now found a better one. A company willing to eat a 12% realised loss to fund a build-out is telling you what it thinks the return on that build-out is, and it is telling you that bitcoin’s expected return, to them, is lower.

Here is where the two filings become one number. Iole’s own disclosed proceeds divided by its own disclosed coin count come to ¥12.11 million per bitcoin. At the dollar-yen rate I read on Sunday morning — 157.52 — that is $76,873 per coin. CoinDesk’s arithmetic on Next Technology comes to $76,976. Two unrelated companies, two regulatory regimes, two accounting conventions, one Friday, and the two implied prices differ by 13 basis points.

That is what a market looks like when the asset is liquid and the filings are honest. It is also the answer to the question the rest of the week kept refusing to give: when the claims are unverifiable, the price is still a value, and you can check it.

The tape

My own reading, Sunday 1 November, 09:32 CET, from the same price feed I use every day:

Bitcoin $77,882.57, up 0.57% on the day. Ether up 0.73%. Nasdaq Composite 25,437.00, down 2.86% at Friday’s close, with the S&P 500 down 1.82% and the Dow down 1.09%. Gold $3,898.90, down 3.29%. Silver down 5.93%. Brent crude $113.43, up 2.37%, with WTI at $104.12 — a $9.31 spread that is itself a chokepoint premium.

Friday’s tape, Sunday’s price - percent change across asset classes

Three days after the largest theft in the industry’s history, bitcoin was the only asset on that list that went up. Gold, the thing you are supposed to buy when institutions fail, fell more than three percent. Silver fell nearly six.

I want to be careful here, because the sessions do not line up: the equity and metals figures are Friday’s closes, while the bitcoin figure is a live weekend quote. Bitcoin trades through the weekend and the others do not, so this is a comparison of a live price to a stale one, not a clean race.

Still. The direction is worth noting, and the reason stocks fell is not the hack. Alphabet’s chief business officer, Philipp Schindler, said on the earnings call 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 summarising product monetising worse than the pages it summarises. That is a business-model problem inside the largest advertising machine ever built, and it landed the same week as a $6.3 billion exchange theft, and the market sold the AI trade and left bitcoin alone.

Which is the uncomfortable symmetry of this whole article. Iole sold bitcoin to buy AI infrastructure. The AI trade then had its worst week in memory. The rotation happened at almost exactly the wrong moment, and it happened because a board needed cash for something it had already promised to build.

The strongest case against all of this

I should make the bear argument properly, because it is better than the bull one this week.

Bitcoin’s most-advertised property — that nobody can freeze it — is exactly what made this week’s outcome worse for the victims. More than $2.4 billion was clawed back on stablecoins because an issuer could pick up a phone and stop it. Not one satoshi of the FBI’s twenty Bitcoin addresses will ever be frozen. The coins are traceable and they are gone, and the thieves will hold them until the addresses go cold. Immutability is a wonderful property when the coins are yours and a terrible one when they are not.

Second: a culture of verification is worth nothing if almost nobody verifies. The FBI published a list that Chainalysis expects to be obsolete within weeks. Binance’s cryptographic proof arrives after the withdrawals reopen, not before. Next Technology does not name its custodians. Twelve DeFi teams outsourced their front end to one registrar and did not notice. Verification is not a property of a ledger; it is a practice, and the practice is rare. The protocol being checkable does not help people who never check.

Third, and the point I find hardest to answer: the advice that actually works — multisig, spread across keys and parties you do not share a failure mode with — is advice almost nobody follows, because it is inconvenient. Nik Bhatia made the honest version of this argument this week: after Coldcard, after Ledger, after Binance, the case for a regulated, insured, boring wrapper is stronger than bitcoiners like to admit. If self-custody fails at the point where a person is tired in a bar at one in the morning, then self-custody as a mass-market default is a claim about human attention, not about cryptography.

What I could not check

Binance’s cryptographic proof does not exist yet; the 3 November date is a promise. Next Technology has not named its custodians or published addresses, and CoinDesk has asked. Chainalysis described its attackers as using “offensive agentic techniques” without publishing mechanics, so I can tell you the phrase and not the technique. The Block’s registrar reporting sat behind anti-bot protection when I tried to pull it, so I have its two summary facts and the rest of the piece is theirs, not mine. And the eltoo channel is a toy.

What I did check, I checked on the ledger. Block 974,104 was mined at 08:24:50 UTC on 1 November and my node is fully synced at that height. The mempool held 36,947 transactions — 12.94 megabytes of people moving value, at a moment when the press was describing the industry as broken. The chain did its job all week. It just could not do anybody’s job but its own.

What to watch

The 3 November proof. Not the announcement — the artifact: a Merkle tree of liabilities, published, with inclusion proofs a user can run, and someone accountable for the total.

Whether the FBI’s address list is an order of magnitude longer in three weeks. If Chainalysis is right about tens of thousands, the list we saw was a photograph of a river.

Whether the eltoo work turns into a BOLT draft. Three iterations, each with real transaction IDs on a test network, is how a proposal stops being an idea. TEMPLATEHASH returning a value rather than a verdict is the kind of change that makes the next ten proposals possible.

And whether the next treasury filing names a custodian. That is the cheapest possible improvement and it costs one line.

The thread I keep coming back to is that the industry spent the week issuing verdicts, and the only party that gave a value was a stranger on a test network building a toy channel for the third time. Ten bytes. The hash on the stack. Everything else was a yes or no, and you could not check any of it.

Sources

Own measurements: cross-asset quotes read 1 November 2026 at 09:32 CET from the price feed I use daily; chain data (block 974,104, mempool 36,947 transactions / 12.94 MB) read from a fully synced node at the same time. Derived figures computed by me and labelled as such: $76,976 and $76,873 per bitcoin from the two corporate disclosures, 57.6x, 12.2%, the 13-basis-point difference, and the >30:1 FBI address ratio.

Write a comment