Coins From Thin Air
The Proxima Mint Exploit, the Freeze and the Reset Question
Somebody turned 200 PROX into 2,000,200 PROX with one transaction, did it again for 5 million, then again for 1.6 million. The bug is fixed, most of the coins are frozen, and by the latest count 3.6 million of them were sold on an exchange before anyone could stop it. A pool operator’s account, with receipts.
On 15 and 16 November 2026 nine ordinary Proxima transactions minted 8,662,267 PROX out of nothing, about 9.5% of the entire supply, by abusing a 16-bit size overflow in the node’s script evaluator. The developer fixed the bug in v1.6.10 and froze 7.4 million of the coins in v1.6.12. About 1.26 million PROX had already been moved on and cannot be frozen. Update 18 November: the developer’s follow-up puts the amount the attacker deposited and sold at an exchange at 3,596,747 PROX, and the exchange’s frozen 2.5 million output is being unfrozen; see Chapter XII. Our pool, sequencer and miners are unaffected. Proxima is, by its own definition, a pre-launch network that may be reset at any time, so a reset is on the table and we look at what it would and would not fix. Below: every transaction ID, how the hole worked, what the freeze does, and the debate it has started.
After this article went up we spoke with the Proxima developer directly. In short: whether to continue with take 1 or reset is “an open question” that will “probably need broader consensus”; they had hoped to carry on with take 1 because “we learned a lot”; and, a correction we accept below, take 1 is not a testnet. Proxima is essentially a one-person protocol team with volunteers around it, and nothing in this article should be read as questioning the protocol work, which held up through a week that would have flattened bigger projects. For our part: we are fine either way. If the network carries on, we keep mining and keep watching the supply; if it were reset with every known bug fixed, we would be back mining from slot one. The numbers below do not change in either case.
18:20 UTC: the developer has since published a full incident report and the decision: take 1 continues, the frozen funds stay frozen, their final fate is undecided. The report, and our reading of it, are in Chapter XI.
Chapter I: Proxima in two minutes, for the newcomers
If you have wandered in from a search engine, a quick orientation. Proxima is an experimental ledger that does away with blocks, mempools and miners-as-validators. Transactions form a DAG (a tangle, in the old IOTA sense) and consensus is reached cooperatively by token holders running sequencers, which keep extending the ledger state with the biggest “coverage”. Time is sliced into slots of 10.24 seconds. The supply target is 1,000,000,000 PROX, of which 6% existed at genesis and 94% is to be mined through a proof-of-work side mechanism: roughly one “mine transit” per slot pays 95 PROX to whoever finds a BLAKE2b hash with enough trailing zero bits. At the time of writing about 91.2 million PROX exist.
We have been involved since the start: we ran the first pool, proxi.suprnova.cc, wrote the first GPU miner, ProxiChorizo, published an FPGA bitstream for it and operate a sequencer, SUPRNV. We rather like the thing. That is precisely why what follows is uncomfortable to write.
Proxima is not a mainnet. Its README calls it a “centralized pre-launch phase” and states that it “can be stopped or reset at any time, without notice, and tokens on it cease to exist at a reset. Nothing is sold and nothing is promised.” Our own Coin Infos page says the same: mine for the exercise. Keep that in mind throughout. It cuts both ways.
Correction (17 November, 10:35 UTC): the first version of this article called the network a testnet. The developer has pointed out, rightly, that it is not: the announced testnet was take 0; what runs today is take 1, the pre-launch network proper, and “otherwise everything is a testnet”. We have corrected the wording throughout. The README’s reset clause quoted above applies to take 1 all the same, which is where our argument rests.
Chapter II: The week everything caught fire
Mid-November was already lively. On 10 November the network switched its proof of work to BLAKE2b (upgrade 6), which we covered in Lemmings With GPUs. On 15 November around 02:40 UTC somebody began broadcasting malformed transactions with impossible far-future timestamps, and nodes across the network started falling over. Releases v1.6.5, v1.6.6 and v1.6.7 followed in quick succession to stop the crashing. Our pool node ran a local fix from 07:01 UTC and the upstream fix an hour and a half later.
Ten minutes after the crash attack began, at 02:51 UTC, slot 234957, a transaction appeared that nobody noticed at the time. It spent a single chain output of 200 PROX and produced outputs worth 2,000,200 PROX. Its total_inflation field reads 2,000,000,000,000 (Proxima counts in millionths). It was not a branch transaction. It was not a mining transaction. It was an ordinary signed transfer that simply declared it had created two million coins, and every node on the network agreed.
We cannot prove the crash attack and the mint were the same actor. The timing, however, is the kind of coincidence that makes a pool operator put the kettle on and open a terminal.
Chapter III: How you mint coins from nothing
Proxima transactions carry small scripts written in EasyFL, a tiny functional language that every node evaluates to decide whether a transaction is valid. Constraints such as “this output needs a signature from key X” or “only branch transactions may create inflation” are EasyFL expressions. A transaction may also carry a free-form script of its own at the end. That is a feature, not a bug: it is how the ledger stays extensible without hard forks.
The bug was in the evaluator’s concat function. It added up the sizes of its inputs and reserved space for the result in a shared memory pool, using a 16-bit size. Ask it to concatenate 65,536 bytes or more and the size wraps around to something tiny. The reservation is tiny; the write is not. The surplus bytes land on top of whatever else lives in that pool.
And what lives in that pool is every other check for the same transaction. Those checks take fresh memory from the pool and assume it is zero-filled. After the overwrite it was filled with the attacker’s bytes, and the rule that forbids an ordinary transaction from creating inflation was evaluated on garbage. Garbage, as it turned out, that said “yes”.
Imagine a bank clerk who checks each cheque on a fresh sheet of paper from a shared pad. Someone hands in a form so long it spills off the sheet and scrawls over the next ten pages of the pad. The clerk then picks up the next sheet, which already has “APPROVED, two million” written on it in someone else’s handwriting, and stamps it. The clerk did nothing wrong by their own rules. The pad was the problem.
You can see the attack script yourself. The raw bytes of each minting transaction, fetched from any public node, end in a long tail of repeated ff and aa blocks: the padding that drives the size past 65,535 and the payload that overwrites the neighbours. The developer’s own regression test, added with the fix in v1.6.10, describes the same mechanism, and the release note says it in two careful words: “stricter allocation bounds”.
The fix itself is small. The developer said on Discord it was two lines, and from the diff that is about right: check the size before reserving, and cap each slice of the pool so that a write past its end can no longer reach shared memory. Small bugs, large consequences: the oldest story in software, now with a token attached.
Chapter IV: The nine transactions
People asked for proof, so we pulled it. Walking the ancestry of every output on the v1.6.12 freeze list back to its root leads to nine minting transactions, all carrying the same script, all signed by throwaway keys, all with is_branch: false and a non-zero total_inflation. Here they are.
| # | Time (UTC) | Slot | Minted PROX | From input of |
|---|---|---|---|---|
| 1 | 15 Nov 02:51 | 234957 | 2,000,000.0 | 200 PROX |
| 2 | 15 Nov 12:53 | 238484 | 10,100.0 | 25 PROX |
| 3 | 15 Nov 12:56 | 238503 | 9,800.0 | 25 PROX |
| 4 | 15 Nov 13:05 | 238556 | 10,463.0 | 25 PROX |
| 5 | 15 Nov 13:19 | 238633 | 11,384.9 | 25 PROX |
| 6 | 15 Nov 14:20 | 238995 | 10,252.8 | 25 PROX |
| 7 | 15 Nov 15:02 | 239238 | 10,266.1 | 25 PROX |
| 8 | 16 Nov 02:01 | 243098 | 5,000,000.0 | 30 PROX |
| 9 | 16 Nov 23:11 | 250542 | 1,600,000.0 | 30 PROX |
| Total | 8,662,266.8 | |||
Amounts from each transaction’s total_inflation field as parsed by our node and cross-checked against the public node at 79.137.70.25.
Notice the shape of it. One big mint, then six little ones of about 10,000 PROX each over three hours, the sort of amounts that look like a healthy afternoon of mining rather than a heist. Then, once the small ones went unremarked, five million in one go, and a further 1.6 million the following night. That is not someone stumbling on a bug. That is someone testing how much they could get away with.
For scale: the whole network mints about 800,000 PROX a day through honest mining (95 PROX per slot, 8,437 slots a day). Nine transactions produced eleven days of the entire network’s output. It took our pool, with a healthy slice of the global hashrate, nearly two weeks to mine 500,000 PROX the boring way.
Transaction IDs and signer keys
Anyone can verify these against any public Proxima node. The parsed view shows total_inflation and the signer; the raw bytes show the attack script.
| # | Transaction ID | Signer (holder ID) |
|---|---|---|
| 1 | 000395cd6a01a1783ddd676404a0de8aa9a22ffdb96eafe8ef270a4a41fd9b0f | 596846c2…fcfb1919 |
| 2 | 0003a3941402c5898443d4abd2e01262760fa67b4d3dc12b95f4305c86260ad5 | d1219278…15452d26 |
| 3 | 0003a3a7140267045d97d618014b9ecef565c24f12a69ac5ddf71b24187060a9 | d1219278…15452d26 |
| 4 | 0003a3dc1402712656e1f17fdb2500ca4a445cb4079244ab755d39aa5c616bdd | f2cbf856…daf8b1b8 |
| 5 | 0003a42914025772562a69f2481d1215ac53e96c05b8ab6addbcda70bd4b4469 | 79c8ec7d…6be709ae |
| 6 | 0003a593140299f53cbb6c95cd7fe7f8c199a00ca1809bc92b5223e1dc032190 | b2dc447b…0e233ab7 |
| 7 | 0003a6861402436b1c44fe2861c240c87015df7caa6b0721934d04741f44b7c3 | e9c0c4b5…0f0c1f45 |
| 8 | 0003b59a2601ac15ecaa951507335af2470af8b50ea57a36f82151aa86f28bef | b6d104d9…a25598d5 |
| 9 | 0003d2ae62010379cf7c8cff0873bf6bb6b7f4f7023ee1b3e80344df2511e9f4 | 20e6f647…242b649b |
Full signer keys: 596846c215f04d8460cae5b6f4302e258bd839dfb2512f0780b3e716fcfb1919, d12192781b46cb78d0088bdc0f5a41886c99711f48267c1ffe84846915452d26, f2cbf8564f100091c4107fc7bfaf8e65a93b56e66991f9d63a35a603daf8b1b8, 79c8ec7dca0bcbc391a45ea518d851181527dfaf14d8d972d5e26e446be709ae, b2dc447b03a3ac2099b2298d052a0c42fc97e9655f4fd30310f7d4c00e233ab7, e9c0c4b533920f5b1f7b18c1453a03c2b85c4f47d1ecc2c3a766d2040f0c1f45, b6d104d93d51a8bca83d70746a1511aee0273c4ba7c9dc036ab01c2aa25598d5, 20e6f647898de967415c39a391ad27771bf68ea2e09869a57c9b85d6242b649b.
To check one yourself, put a transaction ID into either of these endpoints on a public node:
http://79.137.70.25:8001/txapi/v1/get_parsed_transaction?txid=<transaction id>
http://79.137.70.25:8001/txapi/v1/get_txbytes?txid=<transaction id>
One caveat on completeness: we found these nine by tracing the frozen outputs backwards. A mint whose proceeds were entirely spent before the list was drawn up would have no frozen descendant and would not appear in this walk. Nine is therefore a floor, not a ceiling.
Chapter V: The laundering, and the one that got away
Between 16 November 04:08 and 17 November 04:26 UTC the proceeds were peeled off in neat 100,000 PROX slices to fresh keys, then split again across roughly 760 addresses, most holding between 1,000 and 10,000 PROX. Textbook. Spread it thin enough and any freeze list becomes a game of whack-a-mole.
One address stands out: b66e7ba4593dd691eef486bef84c55e72d9f5ca02cc05eb18403585cb0723bee. It funded the chain that produced mint number one, and later forwarded 2,498,049.8 PROX in transaction 0003d9e14402ddc6f04f83429b56f7d32e5c96ec2ea354f557fcdc1a8fef81e0 to a2b09364…4238770a3. Yet its key is not on the freeze list. The v1.6.12 source carries the comment “Custodian holders that also hold funds not on this list: only their fully listed outputs are frozen, their keys are not”, and our reading is that this is a custodian, most likely an exchange hot wallet holding honest customers’ coins alongside the tainted ones. We cannot confirm that and we are not naming anyone. We can say that community members on Discord believe a chunk of the minted coins was sold before anyone noticed, and that the trail going cold at a custodian is consistent with it.
Update 18 November: confirmed by the developer. Both b66e7ba4… and a2b09364… belong to the exchange; the 2,498,049.8 PROX output is the exchange’s hot wallet consolidating attacker deposits that had already been credited and sold on its market. Our transaction read-back shows the same pattern: b66e7ba4… signs, sends 2,498,049.8 to a2b09364… and keeps the change.
| Where the 8.66 million went | PROX | Share of supply |
|---|---|---|
| Frozen by v1.6.12 (781 outputs, 779 addresses) | 7,404,472 | 8.1% |
| Moved on before the list was cut, not frozen | ~1,258,000 | ~1.4% |
| Total minted (nine traced transactions) | 8,662,267 | 9.5% |
Supply 91,220,165 PROX at the time of writing, from get_latest_reliable_branch. Of the frozen outputs, 16 hold about 100,000 PROX each and three hold 484,000, 1.6 million and 2.5 million.
Chapter VI: What the freeze actually is (and isn’t)
Release v1.6.12 is a mandatory upgrade. It ships a file, ledger/blocklist/list.go, with 781 frozen output IDs and 784 frozen signer keys. Every upgraded node refuses to accept a transaction that spends a listed output, and from slot 254500 (around 10:27 UTC on 17 November) refuses any transaction signed by a listed key.
Two things are worth being precise about, because they matter for the argument that follows.
- It is node policy, not a ledger rule. The coins still exist in the state. Every node that matters has simply agreed not to look at them. That works for exactly as long as the upgraded nodes hold the majority of coverage. It is a social agreement encoded in Go, and it has to be carried in every node release from now until the heat death of the universe, or the next reset, whichever comes first.
- It is reactive. The list was drawn up after the fact from what could still be traced. It does not and cannot catch coins that had already left for addresses mixed with honest funds, and it does not catch mints nobody has found yet.
While this article was being written, our supply monitor fired. At 09:22:44 UTC a transaction (254123-69-02d1ef8c086c…) tried to spend the largest frozen output, the 2,498,049.8 PROX sitting at a2b09364…. Both of our nodes refused it with input is frozen, the output is still unspent on every node we can see, and the chain carried on without a hiccup. The freeze works.
Correction, 18 November: we originally read this as the attacker trying the door handle. It almost certainly was not. That output belongs to the exchange’s hot wallet (see Chapter XII), so the 09:22 attempt was most likely the exchange trying to move its own funds, which is exactly the problem the unfreeze now addresses. The developer’s report puts the attacker’s own attempts later, seven of them between 09:31 and 10:28 UTC, starting with the untouched 1.6 million; all were refused.
None of which is a criticism of the developer, who shipped a fix and a freeze within hours of understanding the problem and has promised a full report. Given the week they had, that is good going. Our quarrel is not with the response. It is with what the response leaves behind.
Chapter VII: Our side of the fence
Housekeeping first, since this is also a service notice for our miners.
- The pool node, our sequencer SUPRNV and our separate test node run v1.6.12 since 08:43 UTC on 17 November, well ahead of the slot 254500 activation. v1.6.10 (the actual bug fix) went live on our side at 07:14 UTC.
- None of our keys are frozen. We checked the treasury, the SUPRNV and pool wallets, all 455 pool keys and all 455 miner payout addresses against both lists. Payouts continue as normal.
- One upstream unit test fails on v1.6.12, on the plain tag as well as on our branch: the new check refuses a pulled transaction with a padded signature. We deployed it unchanged, because in a consensus system behaving exactly like everyone else matters more than a green tick.
- We audited every allocation site in the evaluator and Proxima’s built-ins, and fuzzed 300,000 random scripts through a shared pool checking that fresh memory stays zero. The test passes on the fixed evaluator and fails within seconds on the old one. We found no further hole. That is a focused review, not a proof, and anyone can still attach a script to a transaction, so any future evaluator bug remains reachable by anyone.
- A supply monitor now runs every minute on our side. It reads the total supply from our node and sends an urgent alert if supply grows faster than 300 PROX per slot (normal is about 100), if a single slot carries more than 400 PROX of inflation, or if our node refuses a transaction touching frozen funds. The smallest of the nine mints would have tripped it thirty times over. We should have had it on day one. Lesson learned, and we would gently suggest every sequencer operator run something similar.
Because we were watching hashrate, shares, payouts and whether the node was up, like every pool on earth. Nobody watches the supply, because the supply is the one number a ledger is supposed to get right by construction. It was a reasonable assumption and it was wrong. The whole point of running a sequencer is to be one of the parties holding the network honest, and “the supply jumped 2% in one transaction” is exactly the kind of thing that should raise flags with us, not just with the developer.
Chapter VIII: The reset question
Now the awkward bit. Proxima’s README reserves the right to reset the pre-launch network at any time, and a 9.5% inflation bug is about the clearest occasion for using that clause a project will ever get. So the question was always going to come up, and when we raised it on Discord the answer from the developer and most of the community was: the bug is fixed, the coins are frozen, upgrade and move on; the system held up well, a perfect stress test. That is the direction the project is taking, and the developer has since said the question stays open and would need broader consensus to reopen. We are not going to campaign for a reset. We do think it is worth setting out, calmly, what a reset would and would not fix, because the trade-off deserves to be understood rather than shouted about.
What a reset would fix
- The 1.26 million loose coins. Today PROX is worth very little and the unfrozen coins are a rounding error in somebody’s wallet. The whole bet of everyone mining this coin is that one day it is not worth very little. On that day, whoever holds the proceeds of an afternoon with a
concatcall owns about 1.4% of the supply, acquired for 300 PROX of inputs and laundered through 760 addresses. A reset removes that overhang entirely; nothing else does. Plus whatever mints we have not found, because nine is a floor. - The permanent asterisk. Carrying 781 outputs and 784 keys in the node source forever means every exchange, explorer and future audit has to explain that 7.4 million PROX exist in the state but are “not real” by node agreement. A reset deletes the list; continuing means maintaining it in every release.
- The dilution. Every PROX mined from here on is diluted by coins that were never meant to exist. A reset returns the supply to what the emission schedule says it should be.
What a reset would cost
- Everyone’s balances. Honest miners lose weeks of work, exchange buyers lose what they paid for, and the people who sold the minted coins keep whatever they sold them for. The README said tokens could cease to exist, but “you were warned” is cold comfort to someone holding 800,000 PROX they mined fairly.
- The history. Take 1 has survived a BLAKE2b switch, a rollback, a week of crash attacks and this. The developer values that continuity, and from a protocol-engineering point of view they have a point: you learn more from a chain that keeps going than from one you keep restarting.
- The precedent in the other direction. A network that resets after every serious incident teaches holders that nothing on it is ever final. At some point a pre-launch network has to start behaving like the real thing, and the developer’s view is that take 1 already is.
Put side by side: a reset buys a clean supply at the price of everybody’s balances and a month of history; continuing keeps the balances and the history at the price of a 1.4% overhang and a freeze list carried forever. Reasonable people land on different sides of that. We lean towards clean supply, as miners tend to, and we would have been content with a reset done fairly with every known bug fixed. But it is the developer’s network, the README gives them the call either way, and the community that is actually holding the coins has been clear about which price it prefers to pay.
We are not saying the protocol is broken. The cooperative consensus did exactly what it was designed to do throughout a week of crash attacks and a mint exploit, which is genuinely impressive. We are not saying the developer handled it badly; the response was fast and transparent. We are not saying anyone should stop mining, and we are not demanding a reset. We are saying that continuing with take 1 means accepting a known, laundered, un-freezable 1.26 million coin overhang as part of the supply, and that this should be said out loud rather than quietly priced in later. Whatever the project decides, our pool keeps running, keeps watching the supply, and keeps this page up so the record is complete.
Chapter IX: The objections, taken seriously
When the reset option came up on Discord it was, to put it mildly, not popular. Nobody there was defending the attacker; they were defending their balances, which is a different thing and a perfectly human one. The arguments deserve a proper look rather than a shrug, so here they are, steel-manned, with our honest reaction to each.
“A reset would hurt buyers more than the exploit did.”
The exploit has already hurt buyers. Community members say the unfrozen 1.2 million were sold on the exchange on the evening of 16 November, which means the people who bought that evening paid real money for coins minted from nothing, and everyone else’s holdings were diluted by 9.5% in two days. That damage is in the order book now. A reset would return everyone to zero on a network that told them, in writing, that their tokens could cease to exist at any time; continuing leaves the dilution in place but keeps what people have. Which of those hurts more depends on what you hold, which is rather the point.
“The developer did their job. Incidents happen, even on established chains.”
Agreed on both counts, and we said so above. But the reason incidents on established chains are tolerated is that there is no alternative: a mainnet cannot be reset without destroying the thing it is. A pre-launch network can, which is the one advantage it has over a mainnet. Choosing not to use that advantage is a legitimate choice; it just should be made knowingly, not by default.
“It stopped being a testnet the moment it was traded.” / “It never was one.”
This is the strongest objection and we want to be careful with it, not least because the developer makes a sharper version of it: take 1 was never announced as a testnet (that was take 0), and if a traded, mined, running network counts as a testnet then everything is. Fair, and we have corrected our wording above. What has not changed is the README: pre-launch, resettable at any time, tokens cease to exist at a reset, nothing promised. An exchange listing does not amend that, and a third-party order book cannot. If the community believes trading has made this the real chain, that is a coherent position, and continuing with take 1 follows from it. It just means the pre-launch escape clauses are, in practice, retired, and the supply as it now stands, bug-minted overhang included, is the supply the chain launches with.
“They already sold. There is nothing left for them to cash out.”
Then the 1.26 million are now in the hands of people who bought them, which does not make them any less minted from nothing; it just changes who is holding the bag. And the frozen part is not as settled as it sounds: at 09:22 UTC on 17 November someone tried to spend the largest frozen output, and our nodes bounced it. Whoever that is has not gone anywhere.
“At least 7.4 million were frozen. Better than none.”
Much better than none, and credit where due. “Most of the counterfeit notes were seized” is the start of a police press release rather than the end of one, and the freeze list has to be carried in every node release from here on, but as a fix applied within hours by a one-person team it is a good one.
“Stop crying about your tiny investment.”
Fair hit, in a way: everyone in that conversation, us included, was arguing their own book. Someone with 800,000 PROX wants to keep 800,000 PROX, and someone who mined 500,000 the hard way is annoyed that it was out-minted eleven times over by a script. Neither is madness; it is incentives, and worth remembering whenever anyone, us included, says what “should” happen. The README is the only party in that argument with no bags, and it leaves the call to the developer.
Chapter X: Lessons for everyone else
- Watch the supply, not just the hashrate. If you run a node, a sequencer or a pool on any young chain, poll total supply and alert on jumps. It is one number and one cron job. It would have caught this in sixty seconds.
- Scripting languages in consensus are an attack surface. Every function that touches memory in a transaction-validating interpreter is reachable by anyone with a wallet. Bounds-check as if your life depends on it, because the ledger’s does.
- Small mints first. The attacker tested with 10,000-coin mints that looked like mining income before going for millions. Alerts set at “obviously absurd” levels miss the reconnaissance. Set them at “slightly odd”.
- Freeze lists are node policy. If you ever see one in a project you hold, understand that the coins still exist and the freeze lasts only as long as node operators agree to keep it. Price that in.
- Read the README. “Can be reset at any time” is a real clause, not boilerplate. Decide before you mine or buy whether you would be fine with it being used, and equally whether you would be fine with it not being used after an incident like this one.
Chapter XI: The official incident report, and our reading of it
At 18:20 UTC on 17 November the developer posted an incident report on Discord. It is short, numerical and actionable, which is more than most projects manage after an event like this, and we reproduce the substance of it here before adding our own notes. Where our numbers differ from theirs we say so; where theirs are better, we say that too.
What the report says
- Cause. A bug in the EasyFL evaluator let a specially built formula inside a transaction overwrite memory that later checks in the same transaction relied on, so the check of the declared inflation amount passed. “The rules themselves were correct; the program running them was not.” Every node accepted the transactions because every node ran the same faulty evaluator; the report stresses that it “was not an attack on consensus”.
- Amounts. Three transactions minted 8.6M PROX: 2.0M on 15 November 02:55 UTC, 5.0M on 16 November 02:05 UTC and 1.6M on 16 November 23:15 UTC, about 9.5% of supply.
- Fixes. v1.6.10 fixed the evaluator. v1.6.12 froze the funds by node policy. v1.6.13 to v1.6.14p1 made the freeze cheaper to enforce (refusals logged once, honest relaying peers not penalised, no waiting on refused transactions).
- Where the 8.6M are. 7.40M (86%) frozen in 781 outputs with 784 keys that can no longer sign. About 1.10M (13%) reached an exchange before the freeze, mixed with other funds, outside the freeze; the developer understands the exchange is holding the deposits linked to the incident but cannot confirm it from the ledger. About 0.09M (1%) sit in delegations: 63k in the attacker’s own delegation, whose owner key is frozen, and 26k spread thinly over others. About 7k went to sequencers as ordinary fees.
- After the freeze. Between 09:31 and 10:28 UTC on 17 November the attacker tried seven times to move frozen funds, starting with the untouched 1.6M. Every attempt was refused. Sequencers still on older versions accepted the attempts and each one held the network back for about a minute. No attempts since.
- What operators must do. Run v1.6.15 or later; a node without the freeze list accepts the attacker’s attempts and slows everyone down. A node stopped since before 16 November 23:00 UTC must restore from a fresh snapshot, because the fixed evaluator refuses the minting transactions a resuming node would have to replay. Wallets and miners do nothing.
- Decision. “Take 1 continues. We keep the network running and contain the damage, so that it stays as small as possible.” The frozen funds stay frozen; what happens to them has not been decided and will be announced before anything is done.
Our notes, fair and strict
Credit first. Two days from first mint to fix, freeze and a numbered public report, from what is essentially a one-person team, is a good response. The breakdown of where the coins are is more precise than ours, the operational instructions are exactly what operators need, and committing to announce before touching the frozen funds is the right call. We will take the report’s accounting over our own where they differ.
Three transactions or nine? The report counts three mints totalling 8.6M. We found nine totalling 8,662,267 PROX: the same three, plus six mints of roughly 10,000 PROX each on the afternoon of 15 November, between the 2.0M and the 5.0M. They are listed in Chapter IV with transaction IDs; anyone can check them on a public node. About 62,000 PROX is small change next to 8.6M, but those six are the part of the story that shows intent: they look like somebody testing how much would go unnoticed before going for five million. A report that leaves them out reads as if the attacker struck three times out of the blue; the ledger says they rehearsed. We would like to see them acknowledged, and their outputs checked against the freeze list. (The four-minute difference in timestamps is just slot time versus confirmation time and does not matter.)
“Not an attack on consensus.” True in the narrow sense: nobody forked the chain or out-voted the sequencers. But in a cooperative ledger the validation rules are the consensus; every node agreed on a false statement about the supply, faithfully and unanimously. The sentence is accurate and also a little too comfortable. The honest framing is that the consensus layer worked perfectly and replicated a validation bug perfectly, which is exactly what a consensus layer does and exactly why validation code needs the paranoid treatment.
Our 1.26M versus the report’s 1.10M. We said about 1.26 million moved on before the list was cut, from tracing the frozen outputs backwards. The report accounts for 1.10M at the exchange, 0.09M in delegations (of which 63k is effectively stuck, because its owner key is frozen) and 7k in fees, about 1.20M, with the remainder presumably in the six small mints and rounding. Take the report’s figure. The part that matters is unchanged: roughly 1.1 million PROX minted from nothing are now mixed into an exchange’s balances, the ledger cannot see what the exchange does with them, and the developer cannot confirm they are being held. That is a trust relationship with a third party, not a property of the chain.
The frozen funds: decided to be undecided. “What happens to them has not been decided” is honest, and better than a quiet decision. But the options are few and the community should hear them now rather than later: burn them, which restores the supply to what the emission schedule says and is the cleanest outcome; hold them frozen indefinitely, which is the permanent asterisk we described in Chapter VIII; or redistribute them somehow, which we would argue against, because any formula for handing out 7.4 million coins creates more arguments than it settles. We would gently push for a public burn with a date on it.
What the report does not say. When the mints were first noticed, and by whom; whether the far-future-timestamp crash attack that began ten minutes before the first mint is believed to be the same actor; and whether the six small mints are included anywhere. None of these change the response, but a report that is otherwise this precise would be stronger for saying “we don’t know” where it doesn’t.
And the reset? Decided: take 1 continues. We set out in Chapter VIII what a reset would and would not have fixed, and we stand by that analysis, but the call was the developer’s to make and it has been made with reasons given. Our pool continues with it, our nodes follow the release train, and our supply monitor keeps running. If the frozen 7.4 million are ever burned, this chain’s supply will be short exactly the ~1.1 million that got to the exchange plus the small change, and that, a week after launch-grade stress, would be a decent place to end up.
Chapter XII: Follow-up, 18 November: 3.6 million sold, and an unfreeze
A day later the developer posted a follow-up, and at first glance the numbers jump: yesterday’s report said about 1.1 million reached an exchange, today’s says 3.6 million were sold there. We went through it line by line against the ledger and against the earlier figures. Short version: the new numbers are consistent with everything before them. Nothing new was minted and nothing new escaped; what changed is how one large output is classified.
What the follow-up says
- Three attacker accounts deposited 3,596,747 PROX at the exchange. Almost all of it has been sold or spent there.
- The exchange’s output of 2,498,049.8 PROX is being unfrozen. It sits in the exchange’s hot wallet, holds attacker deposits that were sold on its market and therefore now backs what buyers hold; while frozen it blocked the exchange’s larger withdrawals and delegations.
- That leaves about 3.6 million PROX of unplanned inflation, about 4.3% above the supply there would be without the exploit (the exchange deposits plus ~7k PROX of sequencer fees), provided the rest is burned.
- The rest, 4,906,423 PROX in 780 frozen outputs plus the attacker’s delegation of 99,116 PROX, about 5.0 million, will likely be burned. Not decided yet; frozen until then.
- A mandatory release carrying the unfreeze and further fixes is coming; every node operator must update.
Fact-check: does it add up?
| Claim | Check | Verdict |
|---|---|---|
| 3.6M deposited at the exchange vs yesterday’s ~1.1M | 1.10M (unfrozen, at the exchange) + 2,498,049.8 (the exchange’s frozen hot-wallet output) = ~3.60M | Consistent. Yesterday’s 1.1M counted only the unfrozen part; the 2.5M was counted as frozen. |
| 2,498,049.8 PROX is the exchange’s hot wallet | Transaction 0003d9e1… on a public node: signed by b66e7ba4… (not on the key list, “custodian”), pays 2,498,049.8 to a2b09364…, change back to itself, zero inflation | Consistent with a hot-wallet consolidation, and with our Chapter V reading. |
| Remaining frozen: 4,906,423 in 780 outputs | 7,404,472 (our count, 781 outputs) − 2,498,049.8 = 4,906,422.6 | Exact match. |
| ~3.6M = 4.3% above the exploit-free supply | Supply 91.22M at the time, minus 8.66M minted = 82.56M; 3.6M / 82.56M = 4.36% | Correct. |
| Everything adds up to the minted total | 3,596,747 + 4,906,423 + 99,116 + ~7,000 = ~8,609,286 | Between the report’s three-mint 8.60M and our nine-mint 8,662,267. About 53k of our six small test mints is not visible in the breakdown. |
| Attacker’s delegation 99,116 PROX | Yesterday: 63k in the attacker’s own delegation plus ~26k spread over others = ~89k | Not identical; probably reclassified or accrued. Minor, but unexplained. |
So what is true: the 3.6 million figure is not a new loss, it is the honest one. Yesterday’s “7.4 million frozen, 1.1 million out” was arithmetically correct but flattered the outcome, because 2.5 million of the “frozen” coins were already sold to real buyers on an exchange. Freezing them did not hurt the attacker, who had been paid; it hurt the exchange and, through it, the buyers. Unfreezing them is the logical consequence, and we agree with it.
Our notes
The overhang is about three times what we first wrote. We said 1.26 million and about 1.4% of supply. The real figure, if the rest is burned, is about 3.6 million and 4.3%. Everything Chapter VIII said about a permanent overhang applies, at roughly three times the size; it is now sitting in exchange buyers’ balances rather than in an attacker’s wallet, which is better for the chain’s reputation and no different for the supply.
We got the 09:22 attempt wrong. We wrote that the attempt to spend the 2.5 million output showed the attacker trying the door handle. It was almost certainly the exchange trying to move its own hot wallet. We have corrected Chapter VI and apologise for the over-reading.
An unfreeze is a policy act, and it should be documented like one. Freezing and unfreezing outputs by node release is now a precedent in both directions. That is fine as damage control, provided every change to the list is published with the output ID, the amount and the reason, as this one was. Please keep it that way.
Burn, please, with a date. “Will likely be burned” is the right direction. A burn with a fixed slot, announced in advance and verifiable on chain, closes the book; an indefinite “likely” keeps 5 million coins as an open question over every price chart.
Still missing: the six small test mints. The ledger has them; the accounting is about 53k short of our nine-mint total. Small, but a public reconciliation should reconcile.
Nine transactions minted 8,662,267 PROX from inputs totalling 350 coins, using a 16-bit overflow in the script evaluator. The bug is fixed in v1.6.10, 7.4 million PROX are frozen by v1.6.12, and, by the developer’s final accounting, 3.6 million were deposited and sold at an exchange; the exchange’s frozen 2.5 million output is being released and the remaining 5.0 million will likely be burned. Our pool, sequencer and miners are unaffected and now watch the supply every minute. Proxima is a pre-launch network that reserved the right to reset; a reset would have removed the overhang and the freeze list at the cost of everyone’s balances, and the project has decided to carry on. Either way, the best stress test the protocol will ever get is now in the history books, and about 3.6 million PROX of it, some 4.3%, stay in the supply.