One Login, Thirteen Exahash
One Login, Thirteen Exahash
At six in the evening on Friday, 14 November, someone logged into a dashboard and began moving other people’s machines.
The dashboard was Commander, Luxor Technology’s fleet-management software: a web console that lets a mining operator watch a few thousand ASICs and push configuration changes to all of them at once. Pool assignments. Payout addresses. Firmware. The thing it does well is the thing it did that evening, except the person doing it was not the operator. Hashrate from “just over a dozen” customer accounts began pointing at OCEAN, a pool those customers had not chosen. Luxor says its engineers patched the tool within two hours and started restoring the original pool settings (The Block, 15 November).
The machines never stopped. That is the part worth sitting with. The racks kept drawing power, the fans kept spinning, the owners kept paying the electricity bill, and the rewards went somewhere else. CoreX Hosting, a mining host, raised the alarm first and put the affected fleet at roughly 64,000 machines and around 13 exahash. Luxor has not confirmed either figure. But the shape is not in dispute: a single compromised credential in a single piece of software moved a volume of hashrate that, against the network’s 954 EH/s measured on 16 November (mempool.space), is somewhere near 1.4 percent of everything Bitcoin was doing.
The bill for that, as of Monday morning, was three coinbase payouts totalling about 0.243 BTC, roughly $15,700 at prevailing prices, sitting unspent in an address CoinDesk traced through mempool.space. The largest heist of the quarter, measured in dollars, is smaller than a single mid-range car. The lesson is not smaller at all.
The number that kept growing after the attack ended
Here is a detail that separates redirection from theft, and it explains why two respectable outlets published different figures for the same address.
Marty Bent’s account, published on 16 November, counted two payouts totalling approximately 0.162 BTC, about $10,500. CoinDesk, published the same morning, counted three payouts totalling about 0.243 BTC, about $15,700. Both are correct. OCEAN pays out through TIDES, a scheme in which previously submitted work stays eligible for a share of subsequent block rewards over a rolling window (ocean.xyz/docs/tides, as cited by CoinDesk). The redirected machines did not have to find a block themselves. They only had to have been hashing when someone else did, and the money arrives afterwards — after the patch, after the machines were pointed back home, after everybody had gone to bed.
So the amount grew while the incident was being closed. That is a property of the payment design, not of the attacker’s competence, and it tells you something uncomfortable: the accounting of a mining pool can outlive the integrity of the machines feeding it.
OCEAN’s architecture did limit the damage. It is non-custodial: the pool never holds miner funds, it pays the coinbase directly to addresses the miner controls. That rules out the version of this attack where the pool walks away with a balance. It does nothing about the version where the payout address itself is changed upstream, because that change happened in Commander, not in OCEAN. Bent put it exactly right: pool architecture and fleet-management security are separate threat surfaces, and only one of them was tested on Friday.
“A sophisticated AI-enabled vector”
On Monday, Luxor CEO Nick Hansen described the intrusion on X as “a sophisticated AI-enabled vector” and warned that abliterated open-weight models, safety guardrails stripped out and willing to write exploit code, are “running 24/7 against your entire surface, looking for a little crack, which they then use to (almost instantly) expand.” He added that Luxor had put most new feature development on hold while rotating keys and chasing newly disclosed vulnerabilities (CoinDesk, 16 November).
The honest reading is that the technical mechanism has not been published. No post-mortem exists yet. “AI-enabled” is the CEO’s framing of his own incident, and it arrives at a moment when attributing things to AI is cheap and unfalsifiable. It is entirely possible that this was credential stuffing against an admin panel, or a phished session token, or a dependency that nobody watched — all of which are older than large language models and none of which require an abliterated anything.
But the surrounding record is not ambiguous. In July, attackers exploited weak randomness in older Coldcard firmware and drained as much as $114 million from more than 5,000 addresses; Coinkite later said it believed powerful models may have helped find a five-year-old flaw. In August, attackers used BTCPay Server to steal credentials controlling merchants’ Lightning nodes. Later that month, a flood of AI-generated bug reports against Core Lightning turned out to contain real vulnerabilities, forcing emergency guidance to node operators. On 23 October, Bitcoin Core changed its disclosure rules outright, on the argument that AI had made finding code bugs dangerously easy (CoinDesk).
That sequence has a consistent shape, and it is not “AI robs people.” It is “AI collapsed the cost of searching.” The Coldcard flaw sat unnoticed for five years, not because it was well hidden but because looking was expensive. When looking becomes free, everything that was only protected by the expense of looking gets found. Whether the finder then writes the exploit is a separate, much cheaper step — and, as it happens, one that recent research suggests is getting cheaper too.
Six parameters
A paper posted to arXiv on 16 November makes the collapse concrete. Researchers asked whether a general-purpose language model can inject a backdoor into a neural network by editing weights directly, with no training data, no forward passes, no gradients, only reading and rewriting the numbers (arXiv:2611.17395). They could. Across an MLP, a ResNet-18 and a fine-tuned Qwen2.5-0.5B, backdoors went in while clean performance largely held. The MLP needed six parameters. The Qwen model needed 0.02 percent of its weights.
Six parameters. A backdoor that survives a benchmark and fits inside a rounding error.
Two other papers from the same batch point at the same layer from different directions. One shows that memory systems which hard-filter “stale” facts lose accuracy precisely when the resolver doing the classifying is wrong: a controlled intervention swings hard-drop from gains of +0.186 to +0.230 at perfect resolution to losses of -0.176 to -0.181 at the lowest tested reliability, and false-drops hurt significantly more than false-keeps (arXiv:2611.17338). The authors’ recommendation is almost embarrassingly modest — validate on matched data, or, when you cannot, default to keeping things.
The second is an audit of hallucination detectors on the small quantised models people actually ship on-device, rather than the 7B-to-70B research models the detectors were validated against (arXiv:2611.17339). The ranking of four detectors actively inverted between two small models: Spearman rho of -0.80, Kendall tau-b of -0.667. The cheap lexical baseline was statistically indistinguishable from the expensive sampling-based methods. And the standard correctness labeller for one benchmark produced near-chance AUROC of 0.46 to 0.59 — a failure of the label, not of the detector.
Read those three together and the week’s theme stops being about mining. In each case the model was not the failure point. The judgement layer wrapped around the model was: the filter that decides what to delete, the detector that decides what is true, the console that decides which pool to point at. All three are the same object as Commander: an automated verdict standing between a human and reality, trusted because checking was expensive.
The central banker said it in one sentence
The same morning, in Frankfurt, the vice-president of the European Central Bank opened Euro Finance Week with a warning that reads like a description of Luxor’s Friday, scaled up by four orders of magnitude.
Boris Vujčić said AI-related risks to system-wide stability no longer stem mainly from stretched equity valuations. They now come from “the expansion of external funding from equity to debt markets, concentration of exposures, quickening AI adoption and potential cyberattacks.” Then, plainly: “rapid advances in AI models’ cyber capabilities raise risks that have systemic implications.” He flagged the “high concentration of third-party providers of critical services”, ICT in particular, on which euro area banks and intermediaries depend, and noted that a special feature in the Financial Stability Review, due two days later, explores the concentration question further (ECB, 16 November 2026).
That is a central bank naming both halves of the problem in one paragraph: attacks that are now cheap and automated, aimed at a supply chain that has been consolidated for efficiency. Which is the Commander problem. One console, a dozen tenants, one credential.
The other balance sheet
The Austrian reading of what happens next is not mysterious. Credit that is not backed by prior saving does not create demand; it borrows demand from the future and books the difference as growth. When the credit stops, the demand stops, and the assets that were priced on it reprice.
You can watch that mechanism in Nvidia’s filings this month. Jensen Huang resigned as president, CEO and board member effective 12 November, disclosed in a Form 8-K; CFO Colette Kress took over as interim CEO one day before the company’s Q3 FY2027 earnings. Kress, per The Information, had already pushed back on a proposed $250 billion credit backstop for OpenAI that Huang supported; the deal was scaled to roughly $105 billion for an OpenAI Ohio campus signed 17 August 2026. The model Huang built was circular by construction: Nvidia invested in AI companies that spent the money on Nvidia chips. It works exactly as long as the vendor keeps writing the cheques. Reuters then published a review of more than 150 Chinese procurement notices on 13 November finding over 23 billion yuan, about $3.2 billion, in projects specified around export-restricted H100, H200 and H800 parts, while the company’s own 10-K described it as “effectively foreclosed” from that market; IDC put Nvidia’s share of Chinese AI-accelerator shipments at 34 percent in the first half of 2026. Analysts estimate roughly 1.9 million captive custom accelerators were deployed across the hyperscalers in 2026 (a directional estimate, not a disclosed figure), none of it purchasable by anyone outside those four clouds.
None of that is a crime scene. It is a disclosure regime that runs on management’s word, and a financing structure in which the seller is also the lender. Both are the same bet: that the audit will not be run.
What the tape actually says
Now the price, measured rather than narrated. Bitcoin’s August high on Binance was 81,478.87 on 28 August. The low on 15 November was 62,006, a 23.9 percent giveback that took the market back to where the August rally began. From the 260-day high of 87,395.67 on 21 September, the drawdown is 26.1 percent; from the all-time weekly high of 126,199.63 in the week of 6 October 2025, it is 48.9 percent at a spot of 64,469 on 16 November, 09:30 CET (Binance daily and weekly candles, own calculation).

What matters more than the level is the composition of the selling, and three measurements point the same way.
Funding is flat. The last settled eight-hour funding rate on BTCUSDT perpetuals was 0.00072 percent, about 0.79 percent annualised (Binance funding history, 16 November, own calculation). That is not capitulation and it is not a short squeeze in progress. It is a market with no leverage premium left to pay — the froth has already been washed out, or was never rebuilt.
Open interest shrank without a liquidation cascade. Aggregate BTCUSDT open interest peaked at 101,152 BTC on 13 November and stood at 93,524 BTC on 16 November, with notional value falling from $7.23 billion to $6.11 billion over the same three days (Binance futures, own calculation). Positions leaving without a funding spike is the signature of voluntary de-grossing, not forced liquidation.

And the ETF bid withdrew. US spot Bitcoin funds bled a net $2,347 million across the ten trading sessions from 2 to 13 November, nine of them negative, and $2,118 million of that, or 90.2 percent, came out of a single fund, IBIT (Farside Investors). When nine-tenths of an outflow is one product, you are not watching a class of investors change their mind about Bitcoin. You are watching one distribution channel lose its marginal buyer.

Meanwhile the miners did not leave. Network hashrate sat at 954 EH/s on 16 November, with three-day averages running 811, 926 and 1,147 EH/s. Those averages are rising. Difficulty is 134.82 trillion; the last retarget cut it 9.13 percent, and the next one, 1,660 blocks out, is currently estimated at plus 4.73 percent (mempool.space, 16 November). A 9 percent difficulty cut with hashrate climbing is what you see when marginal machines get switched off for a fortnight and better ones absorb the slack. It is not an exodus.
So the honest steel-man for the other side: nothing in the network broke, the difficulty adjustment did its job exactly as designed, and concentration in mining is largely a story about power contracts and uptime, not about ideology. The people who left were the people who had borrowed to be there. That is not a flaw in the observation; it is the observation.
The fix is boring, local, and already written
If the failure mode is an automated verdict that nobody checks, the remedy is a check that does not require permission. It showed up on Delving Bitcoin on 15 November, quietly, with no funding announcement: a Silent Payments scanner for Electrum called spscand, implementing BIP352 (1.1.2), MIT licensed, GPG-signed at v0.1.1, which tracks payments to an sp1 address, follows their spends and reorganises, and reports through a local JSON API and a webhook.
The important sentence in the post is the negative one: it never asks a server about an address, an output or a transaction. Blocks come from your own node, or a blindbit server, or BIP352 filters. Tor works.
That is the entire architectural argument in one tool. Commander failed because the console was the only thing that knew what the machines were doing. A scanner that derives its answers from the chain it already has cannot be redirected by logging in. It can be slower and uglier and less convenient. On the day the dashboard is compromised, it is the only thing still telling the truth.
Consequence
Two failures this week, and the difference between them is accounting.
Luxor’s cost about 0.243 BTC, was detected in hours, patched in two, and will be reimbursed in full. Nobody had to trust a statement about it: the payouts are on chain, unspent, countable by anyone with a node. One compromised credential in one convenience layer cost a company a rounding error and a reputation, and the ledger reconciled anyway.
The AI buildout’s cost has not been booked yet. It sits in $105 billion of vendor guarantees, in 1.9 million accelerators that cannot be sold to anyone outside four clouds, in a $3.2 billion paper trail nobody audited until a newsroom did the document review, and in an ECB speech that uses the word “systemic” about capabilities that made finding bugs cheap. The seller is the lender, the disclosure is unaudited, and the concentration is the business model.
Bitcoin’s advantage here is not that it is faster or cleverer. It is that its failures are visible, priced in public, and settled in a unit whose supply nobody can vote to expand. The machines in those racks kept hashing through the whole incident. The protocol did not need to trust Commander, or Luxor, or OCEAN, or a CEO’s account of what happened on Friday evening — it only needed the blocks to be valid.
Verify, don’t trust is usually said as a slogan. This week it was an accounting statement, and the only party that could produce a clean audit was the one that never asked anyone for permission to check.
Sources
- CoinDesk — Bitcoin miners were hijacked and their BTC rewards redirected to an attacker
- Marty Bent / TFTC — Luxor Commander Hijack Exposes Centralized Fleet Software as a Single Point of Failure
- The Block — Luxor hit by AI-enabled attack that redirected client hashrate, says vulnerability patched
- OCEAN — TIDES payout documentation
- Marty Bent / TFTC — Nvidia’s Vendor-as-Bank Model Ends as Kress Takes Interim CEO Role
- Marty Bent / TFTC — Reuters Found $3.2B in Nvidia China AI Deals the Company Said Didn’t Exist
- ECB — Boris Vujcic, Risks and resilience in the euro area financial system, 16 November 2026
- arXiv — Can LLMs Write Backdoors? An Empirical Security Study of Weight-Space Programming
- arXiv — When Should LLM Agents Forget? Memory Filtering Under Query-Critical Resolver Errors
- arXiv — Does the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device LLMs
- Delving Bitcoin — Silent Payments for Electrum
- James Check / Checkonchain — Red Weekend
- CoinDesk — Bitcoin Core changes security rules as AI makes finding code bugs dangerously easy
- Farside Investors — Bitcoin ETF flow
- mempool.space — difficulty adjustment and hashrate API
- Binance — BTCUSDT spot and futures market data
Write a comment