QWE AI Academy Here’s the part most coverage buried: Claude didn’t crack a password. It found a one-line bug in btcrecover – the open-source recovery tool – where a shared key was being concatenated with the password in the wrong order. The btcrecover tool was concatenating a shared key with the password in an incorrect order; Claude identified the bug, corrected the decryption logic, ran the process, and extracted the private keys in Wallet Import Format. That’s what the May 2026 story is actually about. The viral version says “AI cracked Bitcoin.” The actual version is debugging.
That distinction matters if you want to copy the method. So let’s get into it.
You probably have more recoverable data than you think
The trader, @cprkrn, didn’t just “upload a folder” to Claude. His seed phrase hunt covered two Macs, two external hard drives, an Apple Notes export, iCloud Mail, a Gmail inbox, and X messages – more than 1 GB of data, spread across eight years of devices. The recovery worked because the raw material still existed somewhere, just scattered.
Lost a wallet years ago? Your archive probably looks the same: old laptops gathering dust, a Time Machine backup nobody opens, strings of characters buried in an ancient Notes export. The Claude method is one specific thing – collect every digital crumb from the relevant era, then let an LLM sort which crumb is the one that matters. No magic. Just triage at scale.
Roughly a third of all Bitcoin supply remains held in wallets that have not moved in years, per Glassnode data on dormant coins (as of early 2026). Not all of that is lost – but a meaningful share is sitting behind exactly this kind of “I forgot, but the files are around somewhere” problem.
Why this only works for a specific kind of loss
Claude can’t help if you have nothing. It can help if you have too much. There’s a difference.
- Recoverable with this method: you still have a wallet file (or an old backup of one), and the password exists somewhere – in an old notebook, an email draft, a Notes export.
- Recoverable with this method: you have a seed phrase but it doesn’t open your current wallet file (a sign there’s an older version of the wallet hiding in your backups).
- NOT recoverable: you only remember “part” of a password and want AI to guess the rest. That’s still brute force, and LLMs aren’t password crackers.
- NOT recoverable: you threw the drive away. Claude reads files, it doesn’t read landfills.
Wallet recovery experts told Decrypt (via HotHardware) that Claude’s role was “more of a forensics sorting” than a cryptographic breakthrough. Read that before you upload anything.
The actual setup, step by step
Use Claude with a paid plan – you’ll be uploading a lot of files, and the free tier message cap will kill you mid-investigation. Claude.ai’s web app handles file uploads directly; the Projects feature (available as of mid-2026) lets you persist context across sessions when your dump runs long. Check claude.ai for current plan pricing, which changes periodically.
- Inventory first. Before opening Claude, gather every device and account that existed when you last had access. Cprkrn’s list: two Macs, two external hard drives, an Apple Notes export, iCloud Mail, Gmail, and X messages. Yours probably mirrors that.
- Export, don’t screenshot. Apple Notes → File → Export as PDF. Gmail → Google Takeout. iCloud → download archive. You want searchable text, not images.
- Find every wallet file. On macOS, search for
wallet.dat,.json,keystore,electrum. Don’t open them. Just collect them. - Build a single folder. One folder per device. Date-stamp the parent folder. This is your dump.
- Open a fresh Claude Project. Upload notebooks, exports, and metadata first. Save wallet files for the debugging stage – and read the privacy section below before uploading those at all.
- Prompt with the goal, not the method. “I lost access to a Bitcoin wallet around [year]. I’m uploading every file I still have from that era. Help me identify candidate wallet files, candidate password fragments, and any version mismatches between seed phrases and wallet files.”
Notice step 6. You’re not asking Claude to “hack” anything. You’re asking it to do version reconciliation – the boring part humans are bad at.
The bug Claude actually fixed
This is the technical detail almost every article skipped. Cprkrn found an old mnemonic seed phrase in a notebook, and the HD addresses it generated matched a specific encrypted file – confirming it was the wallet holding the 5 BTC, but the file remained encrypted. The seed matched. The decryption didn’t.
Turns out: early Bitcoin wallets were a different beast. Mnemonic seed phrases generated the HD key tree, but wallets often mixed them with non-HD imported keys – keys that cannot be recovered by the seed phrase alone, stored in a wallet file that requires a separate password. That’s why “I have my seed phrase” wasn’t enough.
# Simplified version of the issue Claude caught
# Wrong order:
decryption_key = sha256(shared_key + password)
# Correct order for this legacy wallet format:
decryption_key = sha256(password + shared_key)The exact bytes differ, but that’s the shape of it. A concatenation order off by one operand. Brute-forcing 3.5 trillion passwords against the wrong key derivation gets you nothing – which is exactly what happened for years.
Pro tip: When a recovery tool fails repeatedly with a password you’re confident in, the problem is rarely the password. It’s usually how the tool derives the decryption key from it. Ask Claude to read the relevant source files of
btcrecover(or whatever tool you’re using) and trace the key derivation path for your specific wallet format. LLMs are genuinely good at this: reading unfamiliar source code faster than you can.
The privacy trap nobody is naming clearly
To make Claude useful, you have to feed it files that contain – or are adjacent to – your private keys. Experts at ZyCrypto warned that uploading encrypted wallet files and password fragments to cloud-based AI creates real privacy and security risks. Worth taking seriously.
Anthropic’s privacy policy states that user-submitted content isn’t used to train models by default, but files still pass through their infrastructure and live in your account history. Two practical hardenings:
- Never upload the unencrypted wallet – only the encrypted file. If decryption succeeds, do it locally on an air-gapped machine using the corrected script Claude wrote, not inside the chat.
- Move funds immediately after recovery to a fresh wallet whose seed Claude has never seen. This is standard operational security for any recovery scenario where sensitive files touched a third-party service.
Cost reality: $15 versus $250 per attempt
Here’s the asymmetry that makes this story actually useful, separate from the AI hype.
| Approach | Cost | Outcome |
|---|---|---|
| Commercial recovery service (per attempt) | ~$250 (per HotHardware) | Failed repeatedly over 11 years |
| GPU brute force on Vast.ai | ~$15 total (per CoinDesk) | Failed (wrong key derivation) |
| Claude (paid plan) | Check current pricing at claude.ai | Solved in a few sessions |
The cheapest tool won. Not because it was smarter than a brute-forcer at guessing – but because it could read the recovery tool’s own code and notice the bug. Specialized recovery services charge per attempt; an LLM subscription lets you iterate until something breaks differently.
Honest limitations
The success case is narrow. According to wallet recovery experts cited by CoinDesk, the process was “more of a forensics sorting” than a cryptographic breakthrough. If your data is genuinely gone, Claude won’t conjure it.
Other real limits worth knowing before you start:
- Context windows fill fast. A 1 GB dump is way past any context limit. You’ll work through it in batches – manifest files, then notebooks, then candidates.
- Claude can hallucinate file paths and password fragments. Always verify a candidate password against the actual encrypted file. Don’t trust the chat output alone.
- Hardware wallet losses are different. If you lost a Ledger or Trezor seed, no LLM helps – there’s no software file to debug.
- Industry estimates put between 2.3 million and 4 million Bitcoin inaccessible – roughly 11% to 19% of maximum supply (per Cointelegraph, as of early 2026) – but most of those losses are seed-phrase or hardware losses, not the kind this method addresses.
FAQ
Can Claude actually break Bitcoin encryption?
No. Bitcoin’s underlying cryptography wasn’t touched in this recovery. Claude found a buggy line in a recovery tool and located a forgotten backup file – both tasks an LLM handles well, neither of which involves breaking cryptography.
What if I only have a seed phrase but no wallet file?
You probably don’t need Claude. Import the seed into any modern wallet (Electrum, Sparrow, BlueWallet) and check whether funds appear. If they don’t, the issue is usually a non-standard derivation path – or, like cprkrn’s case, the old wallet held imported keys outside the seed’s HD tree entirely. The seed physically cannot reach those keys. That’s when you start digging for old wallet files, because only the encrypted file plus the correct password unlocks them.
Is it safe to upload my wallet.dat to Claude?
Encrypted wallet files are useless without the password, so the raw risk is lower than it sounds. But you’re still moving a sensitive artifact through a third party’s servers and into your chat history. If you proceed: treat that Claude account as exposed the moment recovery succeeds. Rotate the account password, enable hardware 2FA, delete the chats, and move the recovered coins to a fresh wallet that Claude has never seen.
Next action: Before doing anything else, run a desktop file search for wallet.dat, keystore, and electrum*.json across every device you owned in your crypto-curious years. If anything shows up, you have a recovery candidate. Then come back and start the inventory.