QWE AI Academy The end result you’re chasing: an old wallet file on your machine, decrypted with a password you already half-remembered, with the BTC swept to a fresh address before anyone else gets near those keys. That’s literally what happened this week.
On May 13, 2026, an X user named @cprkrn used Claude AI to recover an 11-year-old BTC wallet holding 5 BTC – somewhere between $400k and $500k depending on which screenshot you trust. Anthropic’s AI didn’t crack Bitcoin’s encryption; it helped the owner search his own computer for an old wallet file, then a password he’d written down a decade ago did the actual unlocking.
The crypto Twitter takes have been spicy. Some are calling it a quantum-cryptography moment (it isn’t). Others are calling it glorified grep (closer, but still underselling it). This post walks you through what actually happened, what’s reproducible, and where the trap doors are if you try the same thing on your own dusty laptop.
The key takeaway (read this even if you skip the rest)
Claude works as a forensic research assistant, not a password cracker. cprkrn had already tried eight weeks of brute-force on his current Blockchain.com wallet using btcrecover on rented compute, testing about 3.5 trillion combinations with no luck. The breakthrough came when he dumped his whole college computer into Claude and the assistant located an old wallet backup from December 2019 encrypted with a password he already had written down in a notebook.
The reason this matters: bitcoin private keys never change, so an old backup decrypted with an old password gives you the same keys controlling the current funds. You don’t need to break the current wallet. You need to find any historical version of it that opens with something you remember.
Background – why this particular wallet was a puzzle
The setup was nasty. cprkrn bought 5 BTC at a Starbucks in 2015 for about $250 each, stored them in a Blockchain.com wallet protected by three passwords. He still remembered two of them, but at some point he changed the third while stoned and never managed to recall it.
For eleven years, the funds sat untouched. He had a partial mnemonic in an old college notebook, a few candidate passwords, and a graveyard of old hard drives.
That’s the realistic version of “lost crypto.” Most people who lose access aren’t completely empty-handed – they have fragments. As of mid-2026 estimates, industry reports put inaccessible BTC at between 2.3 and 4 million coins, roughly 11-19% of the maximum supply, and a chunk of that is recoverable in theory by someone with the patience to dig.
Method A vs Method B: brute force or file forensics
Two genuinely different strategies exist for recovering old wallets. Most tutorials only cover one. Here’s the honest comparison.
| Approach | Method A: Brute-force the current wallet | Method B: AI-assisted file forensics |
|---|---|---|
| Tool | btcrecover + GPU rental | Claude (or similar long-context LLM) + your old files |
| What you need | The locked wallet file + a guess at password structure | Every old drive, email export, and notebook you can find |
| Cost example | ~$15 of Vast.ai GPU for 3.5 trillion attempts – and still failed | Claude Pro subscription (see current pricing at claude.ai), plus your time |
| Best when | You know the password is short or follows a known pattern | You have fragments scattered across years of digital clutter |
| Outcome here | Failed after 8 weeks | Solved in days |
Method A scales with compute. Method B scales with how much of your digital history you can hand over. The cprkrn case is interesting because he eventually used both – but Method A on the wrong wallet. The whole point is that he was attacking the new wallet when an old backup with a simpler password was sitting in his own files.
How to actually replicate the workflow with Claude
This is beginner-level, but read every step before starting. The exact model used was Claude Opus 4.7, per cprkrn’s own interview with Sherwood News. Sonnet 4.5 also works for most steps and is cheaper – as of May 2026 it carries a 200K-token context window per Anthropic’s official docs.
Step 1: Gather your data – all of it
Pull together every drive, cloud export, and notebook scan that could plausibly contain wallet artifacts. cprkrn’s hunt covered two Macs, two external hard drives, an Apple Notes export, iCloud Mail, a Gmail inbox, and X messages – more than 1 gigabyte of data in total, per Cointelegraph’s reporting. Export your email as .mbox. Take photos of old notebooks and OCR them.
Step 2: Run a local file scan first
Before uploading anything to Claude, find the candidate wallet files yourself. On macOS or Linux:
find ~ -type f ( -name "wallet.dat" -o -name "*.aes.json" -o -name "*.wallet" -o -name "*.key" ) 2>/dev/null
find / -type f -name "*backup*bitcoin*" 2>/dev/nullYou’re looking for wallet.dat (Bitcoin Core), .aes.json (Blockchain.com), Electrum wallet files, or any encrypted blob that looks like it could be one. Move copies (never originals) to a working folder.
Step 3: Hand Claude the context, not the keys
Open Claude and start a project. Upload screenshots of notebooks, text exports, and metadata listings – but think carefully before uploading the encrypted wallet files themselves. Tell Claude exactly what you’re trying to do:
Pro tip: Don’t ask Claude to “crack” anything. Ask it to act as a forensic analyst. The prompt that works is something like: “Here are fragments from my old college computer. I’m looking for any Bitcoin wallet artifacts, backup files dated before [date], password hints, or mnemonic candidates. List every file that could plausibly contain wallet data, with timestamps, and rank by likelihood.”
Step 4: Cross-reference timestamps
This is the move competitors don’t write about. In cprkrn’s case the AI surfaced a wallet backup file from December 2019 that pre-dated his password change – and caught that btcrecover was combining a shared key with the password in the wrong order. The trick is asking Claude to build a timeline: when did you last successfully access funds, when did you change passwords, what files exist from before that date.
Step 5: Decrypt offline, then sweep immediately
Once Claude points you to the right backup and the right password candidate, do the actual decryption on an air-gapped machine. The moment private keys appear, sweep funds to a brand-new wallet on a hardware device. Don’t pause. Don’t celebrate first.
Edge cases the viral threads aren’t telling you about
Here’s where the honest version of the story lives.
The btcrecover bug has wider implications than one recovery story. Claude found that btcrecover was concatenating a shared key with the password in the wrong order – Tom’s Hardware and news.bitcoin.com both confirmed this. If that bug exists in vanilla btcrecover’s handling of legacy Blockchain.com wallets, failed recoveries could have been piling up for years, with users blaming their own memory rather than a tool defect. Check the GitHub issues before committing to a multi-week brute-force run.
The 200K-token context window has a real ceiling in practice. The spec is real, but context fills fast. Heavy users have reported the window saturating almost immediately on large documentation dumps – see GitHub issue #11348 in the anthropics/claude-code repo (filed mid-2025). Feed Claude file listings first, then ask which subset to read. Don’t dump a 1 GB folder in one shot.
Uploading wallet files to a cloud LLM carries real risk. The file is encrypted, so a passive eavesdropper gets nothing without the password. But once Claude helps decrypt it, the private keys exist in a conversation processed on Anthropic’s servers. Those keys touched a third-party system. cprkrn swept funds the same day – treat any wallet you recover this way as compromised by default and plan to do the same before you even run the decryption step.
You may not need Claude at all. If you remember enough of your password structure, btcrecover with a well-crafted tokenlist still beats AI for raw speed on a known wallet format. Claude shines specifically when you don’t know which wallet file to attack – not when you already do.
Realistic expectations
Two paragraphs of honesty before the FAQ.
This story exists because cprkrn had unusually good preservation habits for someone who’d forgotten a password. He kept the old college computer. He kept the notebook. He kept the email exports. If you trashed your 2014 laptop in a 2017 move, no LLM in 2026 will conjure the wallet file out of thin air. Claude is a force multiplier on data you still have – not a substitute for data you don’t.
And the irony of the whole thing: the password was lol420fu**thePOLICE!*:), which the user revealed publicly after sweeping the funds. A complex password he set deliberately, that survived 11 years of forgetting, ultimately recovered by AI rereading old files. There’s a moral in there about security theater versus actual backup hygiene. I’ll let you find it.
FAQ
Can Claude crack a Bitcoin wallet with no backup file?
No. Bitcoin’s elliptic-curve cryptography hasn’t been broken and won’t be by an LLM. Claude only helps if you still have the original encrypted wallet file plus enough context to figure out the password.
Which Claude model should I use for this kind of recovery?
cprkrn used Claude Opus 4.7. For most people, Sonnet 4.5 is the more cost-effective choice – same 200K context window (as of May 2026), lower per-token price, and strong at code-level reasoning like spotting the btcrecover concatenation bug. If your recovery involves running and debugging Python tools – which is half of what made this case work – Claude Code with Sonnet 4.5 is probably what you want. Upgrade to Opus only if Sonnet keeps getting stuck on the same step.
Is it safe to upload my wallet.dat to Claude?
The encrypted file itself isn’t the danger – an attacker reading it gets nothing without the password. The risk is what happens next: the decrypted private keys will exist inside a conversation processed on Anthropic’s servers. Plan your sweep before you start. Have a fresh hardware wallet ready. The moment decryption succeeds, move the funds – same day, not later.
Next action: Open a terminal, run the find command from Step 2 against your oldest drive, and see what comes back. If you get even one hit, you have the start of a recovery project worth pursuing.