Lost Bitcoin Just Got an AI Lifeline — and a New Privacy Problem

X user @cprkrn posted on May 13 that Claude had located an old wallet backup on his college-era computer after eight weeks of failed brute-force attempts using the btcrecover service on a rented Vast.ai GPU — roughly 3.5 trillion password combinations and around $15 of compute, by his own accounting. The wallet had been dormant since April 2015.

“HOLY FUCKING SHIT OMG CLAUDE JUST CRACKED THIS SHIT,” cprkrn posted, vowing to name his next child after Anthropic CEO Dario Amodei, Source: X

What Claude actually did was more mundane — and more important. The model performed forensic triage across thousands of unstructured files on the user’s old machine, surfaced a December 2019 wallet.dat backup, and matched it to a password the user had buried in a college notebook. Because Bitcoin private keys don’t change, decrypting the older backup gave access to the same coins controlled by the user’s locked Blockchain.com wallet. The cryptography held. The metadata cracked.

That distinction matters for the broader pool of lost Bitcoin, which on-chain analysts estimate runs into the millions of coins. With BTC trading near $80,000, even a modest recovery rate would mean tens of billions of dollars potentially re-entering circulation. AI-assisted file forensics now sits plausibly alongside traditional services like Lionsgate Network in the recovery toolkit, though the technique only works where backups and password fragments still exist somewhere — useless for the James Howells landfill case or Stefan Thomas’s locked IronKey.

The privacy tradeoff is the part most of the viral coverage has glossed over. To make this work, cprkrn uploaded the contents of an old personal computer — including encrypted wallet files — to a cloud AI service whose logs may retain that data. For a $395,000 payoff the math may pencil out, but it is precisely the behaviour security researchers have been warning against as AI agents push deeper into crypto infrastructure.

BNC has previously covered Anthropic’s own Mythos research showing that frontier models can autonomously discover and exploit zero-day vulnerabilities in widely-used cryptographic libraries, and the explosion of AI agent activity on crypto wallets through tools like Coinbase’s Payments MCP. The same capability that helped cprkrn reclaim five Bitcoin is, viewed from another angle, exactly what a sophisticated attacker would want to do with someone else’s data.

For now, the takeaway is narrow but real. AI didn’t break Bitcoin. It just made the unloved corner of crypto — wallet archaeology — meaningfully more tractable, with all the upside and downside that implies.