A Bitcoin holder posting on X under the handle cprkrn says he recovered five Bitcoin — worth roughly $395,000 at current prices — after using Anthropic's Claude to untangle a wallet he had been locked out of for years. The story, which spread through crypto and AI circles between May 13 and May 16, has become a high-profile example of large language models being put to work on a very specific, very expensive personal problem.
According to reports from Tom's Hardware, TechRadar and CoinDesk, the owner had changed the wallet's password while high and then forgot the new credentials, leaving the coins stranded. The on-chain movement of the 5 BTC across five transactions on May 13, 2026, lines up with the user's claim that the recovery worked.
How Claude actually helped
Claude did not break Bitcoin's underlying cryptography. Instead, it acted as a methodical assistant for a problem that had defeated brute force alone. The owner uploaded files from an old college computer and asked Claude to help search through them. The model surfaced a wallet.dat backup file dated December 2019 that predated the forgotten password change.
The more important contribution, according to Tom's Hardware, was that Claude identified a bug in how btcrecover — an open-source wallet recovery tool — was combining shared keys and password fragments in its configuration. Before that fix, btcrecover had reportedly tried about 3.5 trillion password combinations without success. With the corrected setup and a mnemonic recovery phrase the owner found in an old notebook, the wallet finally decrypted.
A viral moment for AI-assisted recovery
cprkrn's reaction was less measured than the technical write-ups that followed. "HOLY FUCKING SHIT OMG CLAUDE JUST CRACKED THIS SHIT, THANK YOU @AnthropicAI," he wrote in the post that kicked off coverage in outlets including TechRadar, CoinDesk, Interesting Engineering, Android Headlines and Complex.
Anthropic has not, at the time of writing, issued a public statement on the case. The company has been positioning Claude as a general-purpose assistant for messy, multi-step tasks rather than a single-shot answer engine, and this episode plays squarely into that pitch: rather than producing a clever output, Claude worked through a noisy file system, recognized which artifact mattered and debugged a third-party tool.
What it does and does not prove
The headline number is striking, but the case is narrower than "AI cracks Bitcoin." The recovery only worked because the owner still had fragments of access: the old wallet file, a backup phrase on paper and an approximate idea of the password pattern. Without those, no amount of LLM reasoning would have helped.
It is, however, a clean illustration of where current frontier models shine — synthesizing scattered context, reading unfamiliar configuration files and spotting subtle bugs that human users miss. Security researchers have warned for months that the same capability cuts the other way: tools that help legitimate owners reconstruct lost credentials can also help attackers exploit weak ones. For now, cprkrn's experience is being filed under the upside of that tradeoff.
Reported by Michael Ouroumis.
