Published Updated

Tables and diagrams are easier to read on a larger screen.

Investigation Notes 3 / Lessons

What the drain teaches
an investigator

Nine lessons from where the first reading was wrong or incomplete. Each one is written as something you can apply to the next case, not as a general principle.

Nine lessons

The case, the correction and the method point that came out of it.

LESSON 01 / TIME
Inclusion time is not signing time
  • The type-0x04 EIP-7702 carrier at 20:20:59 held 50 authorization tuples. Only the tuple at nonce 18,914,809 could apply.
  • After the sweep the nonce was 18,914,825, above every tuple. The other 49 could never apply.
  • A block time records when the chain accepted a transaction. It does not record when the signature was made.
  • Use a nonce to test validity and ordering, never to date a signature.
LESSON 02 / ATTRIBUTION
A signature proves use, not a person
The signature proves
  • A valid signature was accepted at that inclusion block.
It does not prove
  • Who signed, or whether a signing service was compromised.
  • Whether key material was copied or an MPC workflow was used.
  • Whether a pre-signed transaction was broadcast later.
The rule
  • Write only what it supports.
LESSON 03 / CONVERGENCE
A shared address is not a shared key
  • Binance 16 needed no delegation and was signed by its own key.
  • It paid the same collection address 24 seconds after Binance 14.
  • Convergence is not proof of one keyholder.
  • Log the shared endpoint, then test the signing paths separately.
LESSON 04 / LISTS
A list is a lead, not control
"are holding or have held assets from the theft, and are operated by or closely connected to" the actors.FBI advisory scope, 31 October 2026
  • The advisory attributed the theft to TraderTraitor and listed 608 ETH, 19 BTC and one SOL address.
  • Its 19 BTC addresses are exactly the case's 19 terminal addresses.
  • A list entry links an address to the theft. It does not prove current control.
LESSON 05 / TOTALS
Keep every total inside one scope
  • At the 1 November cut: frozen four-token 2,271,952,865.67 USD at par; attacker-held 1,002,729,468; unresolved 453,099,600.51; returned 0. Binance's 2.85B frozen or intercepted is a wider official figure.
  • The 2,694,599,999.67 bridge figure is five lines of arithmetic, not a deduplicated total.
  • The withdrawn R4 total added three mixed-scope rows to 3.729B against its own 3.051B subtotal. The 678.2M excess was scope mixing, not overlap.
LESSON 06 / CLAIMS
A claim is not a mechanism
  • "90%, 95%, and 98%" are replenishment from company funds, SAFU, the Futures insurance fund and treasury. The 3 November blog says 98% as of 31 October; the company's own documents are not reconciled. None of it is recovered funds.
  • "Aave unfrozen" was four reserves at LTV 0 and Slope2 5.50/5.50/15.00/30.00 at the cut. By 4 November all four read 10.00/10.00/35.00/50.00. A restriction is not a seizure, and a restored curve is not a return.
  • "THORChain halted" was 10 of 13 inbound chains at the cut, with ETH under a full chain halt and global halt flags false. By 4 November it was 3 of 13. A per-chain halt is not a global halt, and a resumed chain is not a recovery.
  • "Relay $38M bridge" is not corroborated by the Relay index. The named account's 349 transactions have no depository call. One index's absence is not proof.
LESSON 07 / EVENTS
Read the event, then the outpoint
  • One Tron approval carried an Approval event with value 2^256-1 and no Transfer. An approval is not a transfer.
  • Bitcoin principal is 18,270.19988827 BTC across 19 outpoints, all unspent.
  • Dust arrived: 1,111-sat outputs, then 700 and 300 sat. The balance rose to 18,270.20100109 BTC at the cut, and 18,270.20105832 BTC by 4 November, again dust only.
  • Balance change is not principal movement.
LESSON 08 / CORRECTIONS
Correct in public and keep the old line
  • It corrected three swaps to four, and stopped treating inclusion time as signing time.
  • It withdrew an invented 64-record count, a "market maker returned funds" reading and a wrong Arbitrum "not found".
  • It published "five groups, no double counting" for a 3.729B total over a 3.051B subtotal. That was a false total.
  • A correction needs the old claim, the new evidence and the date where the old claim was read.
  • Never delete the old line.
LESSON 09 / THE HUMAN PATH
The signing service is reached through a laptop
  • June 2026. A wallet DevOps engineer downloaded a tampered Homebrew build found through a search ad, to install an AI coding tool. Binance detected it, replaced the device and rotated credentials. The engineer's role and his wallet-team colleagues were already exposed.
  • September 2026. A software engineer was contacted on a legitimate job platform by a fake Project Director. Two interviews followed. The second asked him to download an "assignment" repository onto his work computer. It ran at setup, replaced a legitimate development tool component and opened a channel to an attacker server. The endpoint tools did not detect it.
  • 20 October. The attacker used that computer and an internal AI tool to take control of the wallet DevOps engineer and a wallet software engineer. A container deployment permission token sat on the first engineer's computer.
  • 29 October. The attacker logged into the second engineer's computer with stolen credentials, deployed a container and reached the wallet signing service.
  • Model the human and tooling path as part of the transaction path. A recruiter, a take-home assignment, a package manager and an internal AI tool are all privileged entry points.

Next case

Five actions to carry forward. Each one is a check that would have caught something here.

01Use a nonce to test validity, never to date signing.
02Write what the chain proves, not who signed.
03Put a scope and a source on every total.
04Separate a claim, a control change and a recovery.
05Publish the correction where the old line was read.
Why this page exists. The findings belong in the report and on the landing page. This page is the method: what the case taught, where the first reading failed, and what to check first next time. The correction history is on the notes page, and the evidence is on the evidence page.

Start typing to search the index.

Open the full search page