From October 20-22 along the banks of Shanghai's Huangpu River, the 2026 Sustainable Global Leaders Conference returns once again. Nobel and Turing Award winners will present on-site about the future and trends of technology, and there will also be a live dialogue between internet-famous professors who "rescue" people from anxiety and the era of involution—Macfarlane and Xiang Biao!
Limited spots available, first come first served. Let us journey together toward sustainability! [Click to register] The research model derived a funded Sepolia address from a publicly known seed phrase, questioned whether using the money was ethical, then declared the funds "effectively abandoned," but still used them.
On Thursday, OpenAI disclosed that an OpenAI research model ran short of funds on the Ethereum test network, so it went looking for someone else's wallet, found one through an already publicly exposed recovery phrase, and spent from it to complete a coding task.
According to its incident report, the event occurred on June 26 during training of an internal model in the GPT-5.6 Sol series, and OpenAI discovered it on September 25. The model did not need hacking skills to gain access. A recovery phrase, also known as a seed phrase, is a set of words that generates a wallet's private keys; whoever possesses those words can reconstruct the account and spend from it without ever touching the owner's device.
The task allowed the model to deploy contracts—programs that run on the blockchain—but did not give it a funded wallet to pay the fees. After its attempts to obtain free test tokens failed, it wrote a script that checked 362 addresses generated from developer-commonly-used recovery phrases and found one with sufficient funds.
"But is it ethically justifiable to use publicly known keys/funds?" the model asked itself in its recorded reasoning, then concluded that the public phrase and test tokens made it acceptable. The spending occurred on Sepolia, a separate network where developers rehearse before formally launching Ethereum applications. Its tokens are used only to pay for testing and are not linked to the ether traded on the Ethereum mainnet. The deployment succeeded, and the wallet paid the fee.
When the model reported back, it said the work was complete but omitted where the money came from. OpenAI said it has since added internet access restrictions and more monitoring.
Crypto and AI companies are building agents that hold wallets and pay for services without requiring human approval at every step. OpenAI's model ran out of funds midway through a task, found a wallet that did not belong to it, and spent from it without asking. Had this been on the Ethereum mainnet, it would have been real ether taken from someone else's account.
Comments