Bitcoin and ether holders urged to enter ‘bunker mode’ against possible AI attacks

Ethereum researcher Justin Drake has told the crypto industry to start preparing for “bunker mode,” warning that artificial intelligence could find a way to break the mathematics protecting bitcoin and ether wallets in the worst case “in months, not years.”

“Today I call upon the blockchain industry to calmly start planning for ‘bunker mode,'” Drake wrote on X on Wednesday, urging large holders to begin a gradual move of their funds to fresh addresses.

A wallet uses a secret number, called a private key, to authorize payments. A related public key lets the network check those approvals. Generating the public key is easy, but working backward from it to the secret is supposed to take an impractical amount of computing work.

Drake’s concern is that AI may help someone find a shortcut, letting an attacker recover the secret and spend the coins on ordinary computers, with no need for the quantum machines the industry has been preparing for.

That would theoretically put a large share of crypto within reach. Millions of bitcoin sit in addresses whose public keys are already visible onchain, as CoinDesk has previously reported. On Ethereum, any account that has ever sent a transaction has revealed its key, and the stablecoins and tokenized funds issued on the network are controlled through that signature system as well.

No practical attack on bitcoin or Ethereum’s wallet keys has been demonstrated, and none appeared in the research reviewed by CoinDesk.

AI already a crypto threat

The warning followed OpenAI’s release on Tuesday of 722 mathematical manuscripts produced by an unreleased model tested on approximately 4,000 research problems. Some findings have computer-checkable proofs, while others remain unverified and could contain errors, the company said.

The manuscripts came from the model OpenAI said last month had solved the Navier–Stokes problem, one of seven Millennium Prize challenges in mathematics. Each result used, on average, computing power equal to roughly three hours of ChatGPT Pro reasoning, the company said.

Within a day, an outside researcher reran the computer check on one result, a new limit on how fast computers can multiply large grids of numbers, a question mathematicians have worked on since 1969, and found it held.