Claude Helps Recover a Lost Bitcoin Wallet After More Than a Decade

A bitcoin trader recovered roughly $400,000 in locked funds after enlisting Claude to crack a wallet password lost more than a decade ago.

The approach was methodical rather than novel. Claude helped construct a password-generation strategy based on what the owner could recall about their likely password patterns at the time — character sets, length ranges, structural habits. The model then drove or informed a brute-force process that reportedly tested trillions of candidate passwords against an old wallet backup before arriving at the correct one.

This is not Claude performing cryptographic attacks. Wallet encryption is not broken. What Claude contributes here is the reasoning layer: translating fragmented human memory into a structured search space, reducing the problem from all possible strings to a tractable subset. Without that reduction, the search would have remained computationally infeasible.

For engineers and founders, the takeaway is narrower than the headline implies. LLMs are useful for formalizing fuzzy human knowledge — interview someone about their own past behavior, extract patterns, encode constraints, generate candidates. The same pattern applies to any recovery or search problem where a human expert has partial information they cannot fully articulate on their own.

The limitation is also worth noting. This worked because the owner retained enough contextual memory to meaningfully narrow the search space. It does not generalize to wallets where no such context exists. The cryptography itself was not circumvented.

Practical implication: if you are building tools around secret recovery, account access, or any domain where users have degraded but non-zero memory, LLM-assisted constraint elicitation is a legitimate architecture layer worth considering. Pair it with a targeted search backend and the combination handles problems that neither component solves alone.

Source