⚡️🤖 ICYMI - A developer asked Claude to generate a Bitcoin wallet, sent 12,000 sats to it… and moments later, someone else had already spent them.
He was using AI to automate tests and create multiple wallets, first on the regtest network and then on the live Bitcoin network.
When he sent 0.00012 BTC to one of the generated addresses, the funds were transferred elsewhere almost immediately.
According to him, Claude did not use a cryptographically secure random number generator and simply “predicted” the words that made up the seed.
An LLM like Claude generates plausible text, whereas a secure wallet requires a true source of cryptographic entropy.
If the possible seed phrases are sufficiently predictable, an attacker could theoretically generate them in advance, monitor their addresses, and wait for a user to deposit BTC into them—or use a bot to do so on their behalf.
The developer acknowledges that he was unable to demonstrate exactly how his wallet was compromised, so it is impossible to assert that Claude is directly responsible for the theft.
He was using AI to automate tests and create multiple wallets, first on the regtest network and then on the live Bitcoin network.
When he sent 0.00012 BTC to one of the generated addresses, the funds were transferred elsewhere almost immediately.
According to him, Claude did not use a cryptographically secure random number generator and simply “predicted” the words that made up the seed.
An LLM like Claude generates plausible text, whereas a secure wallet requires a true source of cryptographic entropy.
If the possible seed phrases are sufficiently predictable, an attacker could theoretically generate them in advance, monitor their addresses, and wait for a user to deposit BTC into them—or use a bot to do so on their behalf.
The developer acknowledges that he was unable to demonstrate exactly how his wallet was compromised, so it is impossible to assert that Claude is directly responsible for the theft.
32❤️1👀1🥲1
