Artificial intelligence has picked up a million passwords per second to hack cryptocurrency wallets

A study conducted by NFT platforms NFTEvening and Storible analyzed the capabilities of a neural network based on the Long Short-Term Memory (LSTM) model to recover cryptocurrency passphrases.

Passphrases are keys to digital wallets consisting of 12−24 words, and their security is critical for the safety of crypto assets (ed.).

The experiment tested 85,714,285 combinations of passphrases consisting of 2,048 words from the BIP39 list. The main goal was to determine how quickly and efficiently the LSTM neural network can recover lost or incomplete phrases.

LSTM (Long Short-Term Memory) is a type of recurrent neural network that specializes in working with sequential data, such as text or time series. This architecture has memory cells and three types of control gates (forgetting, input, and output), which allows it to analyze complex patterns and predict missing elements (ed.).

The neural network was trained on a cloud server for 30 days. The average matching speed was 994,051 guesses per second. The recovery time depended on the number of missed words:

• One missed word: 0.02 seconds.
• Two missed words: 29 seconds.
• Three words: 2.28 hours.
• Four words: up to 178 days.

If a part of the cue phrase was known, the neural network could fully recover the 12-word sequence in 8 minutes. If no initial information was available, the matching time increased significantly, and the probability of success became extremely low.

The researchers note that despite the high speed and power of modern artificial intelligence, the number of possible combinations for selecting passphrases remains enormous, making their successful recovery unlikely. Thus, cryptocurrency assets remain safe if passphrases are kept secret.