The NFT-focused platforms NFTEvening and Storible conducted a joint study to explore whether a powerful artificial intelligence model based on Long Short-Term Memory (LSTM) could decrypt cryptocurrency seed phrases. These phrases, consisting of 12 to 24 words, serve as critical keys to access digital wallets.
The study analyzed 85,714,285 combinations of seed phrases derived from the 2048-word BIP39 list. The objective was to assess how swiftly and effectively the neural network could recover lost or incomplete phrases.
For the experiment, the LSTM neural network was trained on a cloud server over the course of 30 days. The model achieved an average guess rate of 994,051 attempts per second, with recovery times evaluated across various scenarios involving one to twelve missing words.
The results revealed that the neural network could restore a single missing word in 0.02 seconds, two words in 29 seconds, and three words in 2.28 hours. However, recovering four words extended to a staggering 178 days.
In cases where partial phrases were already known, the complete and correct 12-word sequence could be reconstructed in as little as 8 minutes. Conversely, when starting from scratch with no initial data, the time required to deduce the seed phrase approaches infinity due to the sheer number of possible combinations.
LSTM (Long Short-Term Memory) is a specialized type of recurrent neural network (RNN) designed for sequential data, such as text or time series.
The researchers concluded that despite the impressive capabilities of modern AI, its potential is fundamentally constrained by the astronomical number of possible combinations, rendering successful seed phrase recovery exceedingly improbable. This underscores the enduring security of cryptocurrencies, provided users keep their seed phrases confidential.
