Introducing Tracer 🚀 https://t.co/1iPtOQBwez @yfarmxtracer
Tracer follows stolen crypto across chains.
🗺️ Trail Map is the free public view of each case. It is in beta: expect errors and incomplete cases. Report bugs. Public cases are published for public interest. Contributions welcome.
First case: the Binance hot-wallet theft, the biggest theft crypto has seen.
🔁 The thief's most effective laundering trick was a compliance refund. A broker on Chainflip refused tainted deposits and refunded each one to a wallet the thief named, nearly always a fresh one. The thief just tried again. 197 retries passed. Every refusal, refund and retry is on chain.
🌀 Then the mixer. It held coins in a coinjoin for a day. On 2 Nov one wallet spent mixer coins in four transactions: 6.41 BTC back into Chainflip, 371,808 USDT out through one Tron wallet by 09:34 UTC.
🌉 A second bridge took $33m out of Tron in ten hours. The wallets it paid, topped up from elsewhere, moved 14,959 ETH into 22 fresh wallets, untouched since 31 Oct.
📦 A lending market played courier: supply under one name, hand the receipt token to a fresh wallet, redeem under another. Six minutes.
🧊 Most of the loot sits still: 18,329 BTC. 537,398 ETH. 378.72m USDT on Tron, blacklisted.
The full trail, step by step, below 🧵👇
We matched money going into Monero with money coming out.
No cryptography was broken. The swap service does its trading on a public exchange, and the amounts line up to the last digit.
The money is from the trail of the theft linked to French bank Crédit Mutuel Arkéa.
Wallets on the trail of the theft linked to French bank Crédit Mutuel Arkéa put $3.65M of ETH into 44 orders to buy Monero on 4 and 5 Nov. For 19 of those orders, the same swap service made a payout 52 to 106 minutes later that fits the order to the last digit. The proceeds of 17 of those payouts, worth $1.34M when sent, ended in one Ethereum wallet.
[No Monero transaction was traced and Monero's cryptography was not broken: both ends of the match sit in public Hyperliquid records.]
1. How a Monero order became public
ZachXBT flagged the suspected theft on 2 Nov. On 3 Nov he put exchange withdrawals since 1 Nov at $133M+. On 4 Nov he named French bank Crédit Mutuel Arkéa as the alleged victim. The same day the bank confirmed fraudulent transfers from its own account and said the money may then have been converted to crypto. On 5 Nov Reuters reported the bank's figure: about €100M stolen.
The $3.65M came from three wallets on a chain of transfers, with one swap of USDC to ETH, from a theft address ZachXBT listed. The chain runs through an older wallet: 88% of what it had received by 03:21 UTC on 2 Nov came from the theft trail.
The swap service, Wagyu, fills its Monero orders on Hyperliquid, a public exchange. For each of the 44 orders it bought XMR1 there, a token it says it redeems for real Monero. Every purchase is on a public ledger: the amount to six decimals, the time to the second.
Wagyu also buys Monero back from customers. On other customers' Monero sales, its own Hyperliquid account sent out XMR1 equal to 99% or 99.5% of the Monero sold. That is public too.
2. The match
19 payouts fit one formula exactly:
XMR1 bought × 0.94 or 0.96, rounded down to 2 decimals, × 0.99 or 0.995
One order: 171.889593 XMR1 bought. × 0.94, rounded down, is 161.57. × 0.99 is 159.9543. 68 minutes after the deposit, Wagyu's account paid out exactly 159.9543 XMR1.
See it on Hyperliquid's public explorer.
The order account, with its deposit at 19:07:44 UTC on 4 Nov and its buys of 169.00 and 3.01 XMR1, which leave 171.889593 after the exchange fee:
https://t.co/4HoNKNqDSP
The payout of 159.9543 XMR1 at 20:15:53 UTC:
https://t.co/tAamceiy1D
All 19 pairs are linked in the reply below.
Our reading: the order delivered 161.57 Monero, and that Monero was sold back to the service with 1% taken off. Other customers' orders show the same kind of gap: on buys, the Monero paid was 2% to 9% below the XMR1 bought, more on bigger orders; on sales, the XMR1 sent out was 0.5% or 1% below the Monero sold.
3. The checks
The search ran from 24 minutes to 6 hours after each order. All 19 fits landed between 52 and 106 minutes.
Reversed, payout before order: 0 of 44 fit.
Same test on 319 other accounts used the same way from 3 to 6 Nov, with one deposit of $5,000 or more and the XMR1 still parked, $23.1M in all: 0 fit, forwards or reversed.
For scale, the same test on every other XMR1 balance of 25 or more, 6,865 accounts, against the service's payouts since July: 5 fit. Three of the five predate the theft. A fit on its own does not say whose order it is. The 44 are tied to the theft trail by where their ETH came from.
The proceeds of 17 of the 19 payouts, worth $1.34M when sent, ended in one Ethereum wallet. It first became active on 4 Nov, after the first order.
The same wallet took more. On 4 and 5 Nov, 26 other payouts by the service, each of 10 XMR1 or more, also led to it. With the 17, that is 43 payouts worth $3.20M when sent. None of the 26 fits any of the 44 orders under the formula, so the evidence for them is weaker. They share the service, the wallet and the dates, but another customer of the service could have theoretically used the same wallet.
We followed every payout of that size from midday on 4 Nov to 02:00 UTC on 6 Nov, 109 in all: 43 led to this wallet. We did not follow the 209 smaller payouts, so 43 is a minimum.
The XMR1 bought for the 44 orders is still parked in the order accounts, as it was on other customers' completed orders.
4. The grade
Order-payout link for the 19 fits: STRONG. Wagyu's order records may show it, or show a refund or another kind of settlement. We have not seen them. Who controls the funding wallets and the receiving wallet is unknown.
Monero was not cracked. The matches come from amounts and timing on a public exchange, at the way in and the way out.
The lesson is opsec, not cryptography: a private chain can still be given away at its edges.
‼️ Someone says they burned 8.9 billion tokens getting Claude Code to make a single strawberry in Blender, the free 3D modelling program.
The creator's own write-up puts it at 2 weeks, 3,958 prompts and 17,453 tool calls.
Malicious code served through Optimizely reached the websites of 178 organisations on 16 November, including Ledger, HSBC, Zoom and CNN. It was built to swap crypto wallet addresses for the attacker's, and from 13:43 UTC it also skimmed card numbers.
https://t.co/DF4edZ5hrM
Binance says an Abu Dhabi court has granted it a $6.5bn freezing order over the crypto stolen in the 29 October hack. The court also told 18 companies holding frozen tokens, including Circle and World Liberty Financial, to keep them frozen.
https://t.co/fVdb6My8AY
CoinShares sold its whole corporate Ethereum treasury in October, plus half its treasury Solana, raising $19.3m to buy back its own shares. Its Nasdaq shares closed at $3.54 on results day, 59% below their first close in April.
https://t.co/h9CvFnuBOD
We gave 4 Claude models the same one-shot prompt: build a detailed 3D pelican riding a bike in Three.js 🚲
Total time · cost · output tokens · $ per 1M in/out:
Fable 5.2: 36m01s · $8.24 · 205k · $8/$40
Opus 5.5: 39m38s · $5.03 · 251k · $4/$20
Sonnet 5.5: 29m19s · $2.33 · 232k · $2/$10
Haiku 5.5: 19m24s · $0.13 · 255k · $0.10/$0.50
About 85% of those tokens were thinking. Haiku used the most tokens of the four and still cost 13 cents. Fable cost 64x more.
Once they got there, all four scenes ran with no errors and no edits, and all four have a visible, textured road. Haiku's pelican hovers a little above its saddle, and the legs are a little weird.
Same 5.2k-character prompt as our Gemini 4 and GPT-6 tests, all through OpenRouter.
We gave 4 OpenAI GPT-6 models the same one-shot prompt: build a detailed 3D pelican riding a bike in Three.js 🚲
Time · cost · output tokens · $ per 1M in/out:
GPT-6.2 Astra: 1m45s · $0.32 · 7.6k · $8/$40
GPT-6.2 Sol: 2m12s · $0.20 · 19.4k · $2/$10
GPT-6.1 Terra: 2m19s · $0.054 · 10.6k · $1/$5
GPT-6 Luna: 2m08s · $0.007 · 14.1k · $0.10/$0.50
Astra, the new flagship, barely thought at all: 234 reasoning tokens, first line of code after 11 seconds. It was the fastest and wrote the least code, yet cost 44x what Luna did.
Sol wrote the most, twice Astra's code, and has the most detailed road: darker, textured and with visible edges.
Terra came in at about a quarter of Sol's cost. Its road is so pale it almost disappears into the white background.
Luna cost less than a cent, but its road faces down, so it's invisible and only the pebbles show. Two of the Gemini 4 models made the same mistake.
All four ran first time with no errors and no edits. Same 5.2k-character prompt as my Gemini 4 test, no follow-ups, all through OpenRouter at each model's default reasoning setting.
We gave 4 OpenAI GPT-6 models the same one-shot prompt: build a detailed 3D pelican riding a bike in Three.js 🚲
Time · cost · output tokens · $ per 1M in/out:
GPT-6.2 Astra: 1m45s · $0.32 · 7.6k · $8/$40
GPT-6.2 Sol: 2m12s · $0.20 · 19.4k · $2/$10
GPT-6.1 Terra: 2m19s · $0.054 · 10.6k · $1/$5
GPT-6 Luna: 2m08s · $0.007 · 14.1k · $0.10/$0.50
Astra, the new flagship, barely thought at all: 234 reasoning tokens, first line of code after 11 seconds. It was the fastest and wrote the least code, yet cost 44x what Luna did.
Sol wrote the most, twice Astra's code, and has the most detailed road: darker, textured and with visible edges.
Terra came in at about a quarter of Sol's cost. Its road is so pale it almost disappears into the white background.
Luna cost less than a cent, but its road faces down, so it's invisible and only the pebbles show. Two of the Gemini 4 models made the same mistake.
All four ran first time with no errors and no edits. Same 5.2k-character prompt as my Gemini 4 test, no follow-ups, all through OpenRouter at each model's default reasoning setting.
We gave 4 Gemini 4 models the same one-shot prompt: build a detailed 3D pelican riding a bike in Three.js 🚲
Time · cost · output tokens · $ per 1M in/out
Carbon: 9m52s · $0.53 · 52.5k · $2/$10
Neon: 5m09s · $0.77 · 51.1k · $3/$15
Hydrogen: 5m08s · $0.31 · 40.5k · $1.5/$7.5
Argon: 5m27s · $0.26 · 25.3k · $2/$10
Carbon looks best. It thought the longest (17.8k reasoning tokens) and didn't start coding until 4m17s in. The result: a textured, winding road with shadows and the most finished pelican.
Neon is the odd one. It's listed as "a fast, mid-priced model", yet it costs 50% more per token than Carbon. It thought less (13k), started coding at 1m54s and wrote the most code, but its road faces down, so it's invisible.
Hydrogen was the only one that worked untouched with everything visible. Argon, Google's flagship reasoning model, reported zero reasoning tokens. It was the cheapest, but its road faces down too.
Same one 5.2k-character prompt, no follow-ups.
We gave 4 Gemini 4 models the same one-shot prompt: build a detailed 3D pelican riding a bike in Three.js 🚲
Time · cost · output tokens · $ per 1M in/out
Carbon: 9m52s · $0.53 · 52.5k · $2/$10
Neon: 5m09s · $0.77 · 51.1k · $3/$15
Hydrogen: 5m08s · $0.31 · 40.5k · $1.5/$7.5
Argon: 5m27s · $0.26 · 25.3k · $2/$10
Carbon looks best. It thought the longest (17.8k reasoning tokens) and didn't start coding until 4m17s in. The result: a textured, winding road with shadows and the most finished pelican.
Neon is the odd one. It's listed as "a fast, mid-priced model", yet it costs 50% more per token than Carbon. It thought less (13k), started coding at 1m54s and wrote the most code, but its road faces down, so it's invisible.
Hydrogen was the only one that worked untouched with everything visible. Argon, Google's flagship reasoning model, reported zero reasoning tokens. It was the cheapest, but its road faces down too.
Same one 5.2k-character prompt, no follow-ups.
gta 6 drops tonight. somewhere in the game, there’s crypto.
first person to find it and reply with the clip wins a year of kraken+ , paid out in USDC 🐙.
your own gameplay only.
kraken+ includes zero trading fees on up to $10k a month of eligible trades and boosted stablecoin rewards.
@rHOODSxyz Yes: the way in and the way out stay public and services still keep records, but the 16-coin lists that let you test a timing guess are gone on new spends.
This week we traced funds from the Gnosis Safe theft to the edge of Monero, then found a pattern inside it.
Monero normally hides each spent coin in a list of 16: the real one plus 15 decoys, with nothing marking which is which.
By timing, we picked out the Monero payouts we think the stolen funds were swapped into. Then we found one transaction where 10 of the 11 lists include a coin from one of those payouts.
Chance? We ran the count 10,000 more times with random coins from the same blocks in place of ours. The most chance ever matched in one transaction was 5. Ours matched 10.
Monero is testing an upgrade that removes the list. 🧵⤵️
@IsFinkle@sillywuh It is in the initial thread: "25 lots sent through its Monero bridge". The ring data is posted too, with every hash, for anyone who wants to check it and not just shout.
We linked $4.06M of the Gnosis Safe theft inside Monero.
28 payouts located. Ring analysis leads all 28 into one group of 14 transactions. Across 9,634 transactions scored, two hit 10 and 6. No other got past 2.
Still watching the exits.
Monero is built to hide who pays whom and how much. We now hold ring-based leads that carry about $4.06M of the Gnosis Safe theft onto the Monero blockchain and one transaction further.
The money went into a Monero swap service in 42 pieces. We located the Monero payout for 28 of them, and ring evidence leads all 28 into one group of 14 Monero transactions.
WHAT HAPPENED
• On 16 November, $6.43M left a Gnosis Safe in one transaction.
• Within 28 hours about 98% of it had gone into one Monero swap service, split into 42 pieces: 17 swap orders and 25 lots sent through its Monero bridge.
WHAT WE FOUND
For each of those 42 pieces, the service pays the buyer in Monero with one Monero transaction. We call that transaction the payout. Finding it is the step from the public side into Monero.
• Payouts found: 28 of the 42. That covers 23 of the 25 bridge lots and 5 of the 17 orders.
• Where they lead: ring evidence links all 28 payouts onward into one group of 14 Monero transactions. Monero bought by three of the thief's wallets meets in that group, which reads as one holder gathering the coins.
• How much that is: about 7,996 XMR ($4.06M) is linked this way.
• Candidate links: the other 12 orders, about 3,380 XMR ($1.72M), point to the same group. They match as a batch and other customers' payouts sit in the same blocks, so we count them as candidates.
WHAT RING EVIDENCE MEANS
Every Monero spend lists the real coin hidden among 15 decoys. When later transactions list the payouts we located far more often than chance allows, that pattern is the link.
WHERE IT STANDS (22:10 UTC, 17 November)
• No link from the group to a door out of Monero is seen.
• The group's open outputs are under watch.
Monero hides amounts, so each XMR figure is an estimate from the size of the lot or order.
How we did it, and what it lets us test next, is in the next post in this thread 🧵⤵️
@whatmonero@sillywuh Read the first post again: the money went into a Monero swap service, 25 lots through its Monero bridge. XMR1 is that bridge token. Nothing changed. The payouts are real Monero on the Monero chain, and the hashes are posted in this thread.
Fair ask. Here is the data and the methods we can share.
The two transactions:
https://t.co/DwR1kLX3H7
https://t.co/NCs9vbK5xA
One send followed end to end:
1. 17 Nov, 07:39:53 UTC. An account funded from the stolen money sends 301.133382 XMR1 to a one-off address, which passes it to the swap service 33 seconds later.
https://t.co/zqkXgzPWYn
2. The first Monero block stamped after that is 3,808,243. It holds this two-output transaction, which we read as the payout:
https://t.co/kSijCB8X3c
3. One of its outputs (key 308f5fb0…bc21) is listed 20 minutes later in the first transaction above, under key image 01, ring member 14.
4. Its other output (key b46ba003…676c) is listed under key image 04, ring member 15, in a later transaction we put on the service's side:
https://t.co/Je4taHOMsm
One listing could be a decoy. So we counted the pattern.
How we counted:
1. Start with the 25 XMR1 sends from the three accounts funded from the stolen money, and the time each reached the swap service on Hyperliquid's ledger.
2. For each, take the first Monero block stamped after that time and the next two. In those blocks, find every two-output transaction with one output listed in a ring of the seven transactions we put on the service's side. Its other output is a candidate. That gives 88 candidates from 87 transactions.
3. Check every transaction with 2 to 16 inputs in blocks 3,808,192 to 3,808,620, leaving out miner rewards and those seven: 10,157 transactions. A transaction's score is the most rings we can match at once, each to a different candidate transaction.
Scores: 10 for one, 6 for another, and nothing else above 2. The 10 line up with all ten sends the first account had made by then. The 6 line up with all six of a second account.
The luck test: swap each candidate for a random output from a random two-output transaction in the same block, and score everything again. We did that 10,000 times. The best score was 5, once.
Without the time windows, two other transactions score 12 and 8, and we don't know whose they are. Other accounts were sending XMR1 to the same service that day: from 05:16 to 15:25 UTC, 29 sends from five other accounts reached it. That's why the count is tied to the sends on Hyperliquid.
What we keep back, and why: the number of extra moves at which the pattern stopped standing out in our test, and how we chose the seven service-side transactions. Either could help the holder get clear of this check.
What this covers is the gathering inside Monero. The coins haven't left as far as we can see.