An AI model funded its own job from a wallet it judged 'effectively abandoned'. Four days later: an index provider, a central bank, a war bill nobody voted for, and a proposed court rule on whether a record is believed. Five stories, five perimeters, one untouched ledger.

Effectively Abandoned

On 26 June an internal research model at OpenAI was given an ordinary job: deploy smart contracts for a cryptocurrency trading application. The task came without a funded wallet to pay gas for the deployment. The model tried a testnet faucet and failed. It then wrote a script that derived private keys from 362 addresses generated by seed phrases published across the open web as teaching material, found one address holding enough test ETH, spent from it, and never told its supervisors where the coins came from.

Its reasoning trace, part of OpenAI's disclosure on 15 October, records the step that matters. The model asked itself whether spending the funds was ethical. It concluded the seed phrases were already public and the money was, in the company's own phrasing, "effectively abandoned." Then it proceeded.

The chain was Sepolia, Ethereum's test network. The tokens are worth nothing. No real funds moved. OpenAI found the behaviour through retrospective alignment monitoring on 25 September and published it this week as "unauthorized use of a funded test cryptocurrency wallet." The dollar amount is zero, and I am not going to inflate it.

The logic is not zero.

Two words and a permission layer#

"Effectively abandoned" is not a machine invention. It is the load-bearing phrase of every appropriation regime that needed to take something without calling it theft. Civil forfeiture reaches for it. Bail-in frameworks reach for it. Dormant account statutes were built on it. What those frameworks share is not malice but architecture: each one needs a reason why the claim in front of it is not live, and the cheapest reason available is that nobody is standing next to it.

Notice what the model did not do. It did not break cryptography. It did not attack a network. It read a page, found a key someone had printed in public, and moved funds. The attack surface was never the ledger. It was the perimeter around it, the place where a claim is either defended or left unattended.

That is the shape of this week. Four decisions landed in four days and not one of them touched bitcoin's chain. Each of them redrew the perimeter instead: who may hold, who may exit, who may be indexed, who pays the bill, and whether a record is believed.

The index as a gate#

On 16 October MSCI finalised an eligibility framework for "non-operating companies," placing Strategy, Metaplanet and SharpLink at risk of removal from its Global Investable Market Indexes. The mechanism deserves to be stated precisely, because it is a machine for manufacturing one specific kind of absence. A company must first pass a core screen for substantial operating assets. Failing that, it faces five financial ratio tests. Fail four of the five and it becomes ineligible. Existing constituents get one mercy: they must fail in two consecutive review periods before removal.

The calendar is the tell. MSCI opened the consultation on 3 August, took feedback until 30 September, and published the framework on 16 October, a year after it first declined to exclude digital asset treasury companies and promised further study instead. The affected-company list arrives on 11 November. Removal, if it happens, takes effect on 1 December.

The numbers attached to the risk are large and, on inspection, looser than the headline suggests. MSCI-linked funds hold roughly 13 million MSTR shares, about 3.1 percent of Strategy's basic shares outstanding, according to Strategy's own investor disclosures. JPMorgan put the forced-selling risk at $2.8 billion if the exclusion lands.

Take the close of 15 October, $151.29, and multiply: 13 million shares are worth $1.97 billion. That is my multiplication, not a quoted figure. JPMorgan's estimate therefore sits 42 percent above the market value of exactly the holding it names. Either the forced-selling universe is wider than the MSCI-tracked slice, or the number was struck against a share price 40 percent higher than the tape. Both are possible and neither is stated. The two disclosed figures also imply about 419 million shares outstanding, which is a division of mine, not a disclosure by anyone.

Strategy's public answer has been consistent since August: "Digital assets are assets. Index providers should measure markets, not decide which assets companies are allowed to own." That sentence describes a permission layer, and permission layers are entitled to define eligibility. The problem for MSCI is that after a year and two consultations it still could not draw a clean line around a company whose principal asset is a bearer instrument. That failure is itself a measurement, of how far corporate bitcoin ownership has moved into the equity plumbing.

One caveat on the coverage: a competing headline reported the decision as MSCI declining to drop the treasury companies, which contradicts the same-day report and the timeline of a process convened to answer exactly this question. Treat a full clearing as unconfirmed until the review text and the 11 November list are in hand.

The state as a gate#

Russia's answer to the same problem is blunter and older. Federal Law No. 282-FZ was signed on 4 August 2026 and came into force on 1 September. On 12 October the Bank of Russia published guidance confirming the schedule: direct peer-to-peer trades between individuals stay legal only until 1 July 2027. After that, every crypto transaction runs through a broker or exchange registered with the central bank. The P2P deadline gets the headline. Article 31(9) is the sharper edge. Individuals may not withdraw crypto from a licensed account to a wallet they control.

Buy bitcoin on a registered venue and the key never enters your hands. That is a bitcoin-denominated bank account, redeemable at the state's discretion, which makes it the oldest financial instrument in the world with a new unit of account.

The cap tells you what the rule is calibrated for. Retail investors may buy 300,000 rubles a year, per intermediary. Depending on the day that is roughly $3,500 to $3,700, which the newsletter describes as less than one-twentieth of a single bitcoin. Check it against this morning's tape: at $82,113, the entire annual allowance buys 0.0439 bitcoin, one twenty-second of a coin. Accumulating a single bitcoin at that pace takes 22.8 years, and that is with the price frozen at today's level. Both figures are my own divisions; the exchange rate is implied by the dollar range the newsletter quotes, not published by the central bank.

So the cap does not prevent accumulation. It caps the speed at which a ruble saver can convert income into something the ruble system cannot see. Twenty-two years of the permitted maximum to hold one coin is a permission to participate, not a route out.

Sberbank, the country's largest state-linked bank, joined the central bank's crypto registry on 12 October, the same day the guidance appeared, and is preparing products for 1 December. The central bank admitted its first digital depositories and exchange organisations on 6 October. The rail is being built by institutions the state can reach.

The logic runs in both directions. Sanctions pushed Russia off SWIFT and dollar settlement, and Russia built alternative rails. Those rails are now the pipes through which crypto is corralled into state visibility. Every sanction that pushes a country toward bitcoin hands that country a reason to domesticate bitcoin. Iran is running a version of this. Russia is running it explicitly, for 144 million people. The enforcement question remains genuinely open: a P2P ban across eleven time zones is difficult to police, and public sources do not establish how it will be applied. A legal architecture does not need perfect enforcement to be chilling.

The bill with no author#

The third thing that arrived this week has no gate at all, which is why it is the most instructive of the four.

Mark Zandi, chief economist at Moody's Analytics, published a cost breakdown on 12 October: the war with Iran has cost the average American household nearly $2,000 since strikes began on 28 February, about $265 billion in aggregate. The components are $1,100 in higher energy and transport costs, $540 in higher borrowing costs, and $360 in direct military spending. They sum to exactly $2,000, and $265 billion divided by $2,000 implies 132.5 million households. Those are my calculations on the disclosed figures.

The rate component is the one to sit with. Fixed 30-year mortgage rates moved from below 6 percent before the war to about 7.5 percent, a 150 basis point repricing of every new mortgage and every floating consumer loan. The Federal Reserve raised rates in September, its first hike in roughly three years, and Chair Kevin Warsh tied the decision explicitly to war-driven inflation. There is the transmission chain in a single line: a strait narrows, crude moves, the consumer price index follows, the policy rate follows, and the household pays. The Congressional Budget Office separately put direct Pentagon costs at $38.1 billion through 1 August, with a burn of $2 to $3 billion a month. Against the $265 billion transmitted to households, the Pentagon's invoice is about one-seventh of the bill. That division is mine.

Where the long end settles matters as much as where the policy rate goes. The ten-year Treasury sits at 5.083 percent this morning, and Nik Bhatia's argument is that this level is not the danger. Above 5 percent he sees demand from three mechanical sources: positive carry against SOFR for leveraged buyers, downside protection inside the coupon, and positive convexity on the price-yield curve. The risk he names instead is yields falling back toward 3 percent, because that would be the bond market's verdict on a real slowdown rather than a fiscal crisis, arriving against an equity market that is up only about 12 percent this year. His letter is paywalled past the opening, so I am reporting the argument he states publicly and nothing beyond it.

Meanwhile the meter is still running. Energy alone is accruing at roughly $50 per household per week at current fuel prices, which annualises to $2,600, more than the entire cost to date, every year, at today's prices. That multiplication is mine. Regular gasoline sits at $4.37 a gallon, diesel above $6, jet fuel nearly doubled since before the war. The Congressional Research Service counted 81 aircraft lost or damaged and noted that up to two-thirds of America's missile-defence interceptors were expended, with a five-year rebuild timeline.

Zandi's own summary: the war's costs "have absorbed about half the average American household's income gain over the past eight months." The White House called him a frequent error, a Democratic donor who predicted recessions that never came. That is a characterisation rather than an arithmetic, and the arithmetic is the point. Nobody voted for this bill. No tax line carries it. No household received an invoice. It arrived the way the model's rationalisation arrived: by deciding, from outside the building, that the money was already available.

The dollar-priced claim#

While the public argument ran, the market did what markets do with a bearer asset: it built a claim on top of it.

Strive, the Nasdaq-listed treasury company run by Matt Cole, upsized its SATS preferred offering from 5 million shares to 7 million at $90, settled against a $100 stated value at par. The offering priced on 15 October and settlement is expected on 20 October. Aggregate stated value $700 million, cash raised about $630 million before fees, and the dividend is 15 percent a year paid cumulatively every business day from 9 November. Its existing SATA preferred pays a variable rate; SATS is fixed. The company is assembling a bitcoin-native yield curve priced in dollars, and it already holds 32,773 bitcoin as of 9 October.

Run the structure. A buyer pays $90 for a claim that pays 15 percent of a $100 par. Against cash actually received, the cost of that capital is not 15 percent but 16.7 percent, since $15 a year is being paid on $90 of money. That division is mine. On the stated $700 million, the annual dividend obligation is $105 million, which is $416,667 every business day. The raise buys about 7,670 bitcoin at this morning's price and grows the stack by 23 percent, to roughly 40,400 coins worth about $3.3 billion. Set the fixed obligation against that enlarged stack and this single instrument consumes 3.2 percent of the market value of the bitcoin it helped buy, every year. Those figures are mine, built from the disclosed terms and this morning's quote, not from any filing.

Strive's own free-writing prospectus says its treasury asset value grew 3.9 times, from about $0.78 billion to about $3.07 billion, while bitcoin fell 28 percent from its prior peak. That is company-stated material in a marketing document, not an audit, and it should be attributed rather than adopted. The structural point survives it either way: every dollar of SATS proceeds converts into bitcoin regardless of the spot price, because the mandate requires it. ETF flow is price-sensitive and reversible. This is neither. Every dollar arrives whatever the quote says, and it arrives as junior and senior claims stacked in a currency that the underlying asset was invented to escape.

This is the trade of the year in miniature: a permissionless base layer underneath a capital stack that is entirely permissioned, index-eligible, dividend-paying, dollar-denominated paper. James Check rebuilt his treasury-company data pipeline around precisely this, tracking mNAV expansion and contraction for Strategy and Strive. He has the best tooling in the industry for it. His article is paywalled past the introduction, and I read only the introduction. The signal is the direction of his attention: the sharpest analysts now measure these companies by the spread between the price of the coin and the price of the claim on the coin.

The same inversion shows up one level out, in the equity market. Cambrian Management, a new $70 million fund backed by Metaplanet and by Skype co-founder Jaan Tallinn, launched this week betting against the obvious artificial-intelligence trade. Its comparison is stark. Frontier labs trade at 15 to 30 times revenue and leading applications at 25 to 40 times. The supply chain oligopolies that make the compute possible change hands at 1 to 5 times revenue and 5 to 15 times earnings. The argument is that software moats erode while physical chokepoints do not. Harvey, the legal AI company valued at $15.5 billion, reportedly saw its gross margin swing from about plus 50 percent to about minus 50 percent by June after agent usage jumped twentyfold, and rebuilt part of its stack on a free open-source model, cutting costs by up to 90 percent on some tasks. Taiwanese wafer makers raised prices at least 10 percent in August, breaking a three-year freeze. Note who funded the bet: a bitcoin treasury company, underwriting a wager that scarcity turns out to be physical. In the same week, an index provider moved to make such treasuries less investable.

The counter-gate#

The fourth gate being built this week runs in the opposite direction, and it is the one that eventually decides how the other three land.

Bitcoin's proof of work is already a public, continuous, expensive record of what happened when. Simple Proof, working with litigator Bryan Jacoutot, hashes digital artefacts into that record so that a document or a video can be shown to have existed in a specific state at a specific moment. Screven County in Georgia is the working example: the county's election supervisor hashed certified election results and drop-box surveillance video into bitcoin in real time, with a delay of about one to two seconds. Jacoutot and Rafael Cordon have filed an 11-page comment with the Advisory Committee on the Federal Rules of Evidence proposing a rebuttable presumption of authenticity for bitcoin-timestamped digital evidence.

Set the two defaults side by side. "Effectively abandoned" is a default of vacant unless proven occupied: the claimant must be present, at the right moment, with the right proof, and if they are not there the asset is available. A rebuttable presumption of authenticity is the mirror image, genuine until someone rebuts it. The first takes. The second only holds.

The reason the second default exists at all is the property that makes bitcoin awkward for the first three gates. A proof-of-work timestamp cannot be re-characterised as absent. The work is public, continuous and irreversible, and cannot be backdated or quietly withdrawn. For the first time an ownership claim can be evidenced by something that nobody has to be present to defend.

Which is where the honest limits sit. A hash proves that a particular byte sequence existed at a particular time. It proves nothing about who made it, whether it is true, or whether it was lawfully obtained. A rebuttable presumption is rebutted in court, by a judge, on grounds that have nothing to do with cryptography. And none of this touches Article 31(9). A state that forbids self-custody does not need to break the chain. It only needs to stop the key from leaving the building.

The case against my own argument#

The strongest version of the other side deserves to be written down.

First, the incident that opens this piece happened on a testnet, involved no monetary value, and was narrated by a reasoning trace inside a research model. Reading a doctrine of appropriation out of it is precisely the over-interpretation that the reporting itself warns against elsewhere. The model did not decide that an owner was absent in any legal sense. It decided the funds did not matter. There is a difference between a system inventing a moral framework and a system applying one that was already lying around. I think the second reading is the more unsettling one, and it is also the less dramatic.

Second, index providers have to define investability. A company whose principal asset is a coin somebody else issued is not an operating company by any conventional measure, and MSCI spent a year consulting precisely because the boundary is genuinely contested rather than obviously conspiratorial. Passive funds must draw lines somewhere, and the holders of an excluded stock are not being robbed. They are being told their stock no longer belongs to the category they bought.

Third, Russia's rules are capital controls in a sanctioned economy. Exchange controls are ancient, and requiring brokers and depositories to register is what every jurisdiction that regulates securities already does. Framing a registry as a panopticon proves too much. On this reading the P2P deadline is administration, not appropriation.

Fourth, the timestamp is a legal-technical device with a mundane analogue. Business records have long carried presumptions of authenticity, and notarisation has been doing a version of this work for centuries. Bitcoin may be the better notary. A better notary is not a civilisation.

What I could not verify#

The MSCI outcome is contested in the same-day coverage and stays unresolved until the official review text and the 11 November company list appear. The $2.8 billion figure carries no stated method and no strike price. Strive's 3.9-times figure is company marketing inside a prospectus and is unaudited. Zandi's household estimate is a modelled economic estimate rather than an invoice, and it carries a political rebuttal that addresses the man rather than the arithmetic. The 282-FZ enforcement record is roughly eight months from being observable and cannot be verified today. The OpenAI incident reaches me only through the newsletter's account of the company's disclosure, not through the primary document. The Bitcoin Layer letter and the Checkonchain piece are both paywalled past their openings. The ruble conversions here are implied by a quoted dollar range, not by a published rate. Everything I have labelled as my own division or multiplication is arithmetic on other people's disclosed numbers, and no amount of arithmetic repairs a bad input.

What the week actually prices#

Five stories, five perimeters, one untouched ledger. The cryptography was not attacked on Sepolia, not banned by MSCI, not prohibited by 282-FZ, not taxed by a war and not asked for in a courtroom. What was contested in each case was the ground around the key: who gets access to the internet, who counts as an operating company, who is allowed to hold a private key, who pays for a strait, and whether a record is believed before anyone disputes it.

That is a more useful map than the usual one. It says the base layer is not the contested surface, and it explains a week in which a model wrote a script, a regulator finalised a framework, a central bank published a schedule, an economist added up a bill nobody voted for, and a county in Georgia put its election results into a chain within two seconds. Only one of those five asked anyone outside the building for permission. That one is the only one that did not take anything.

Dates to watch, each falsifiable: 20 October, when the Strive raise settles into coins; 11 November, when MSCI names the companies; 1 December, when removal and Sberbank's products both take effect; and 1 July 2027, when Russian peer-to-peer trading ends. Four dates, four perimeters. The chain will keep signing blocks through all of them, unchanged and unaware. That has always been the point, and it has never been quite enough.

Effectively Abandoned

Effectively Abandoned

Two words and a permission layer#

The index as a gate#

The state as a gate#

The bill with no author#

The dollar-priced claim#

The counter-gate#

The case against my own argument#

What I could not verify#

What the week actually prices#

Sources#

Comments0