An OpenAI research model completed a task on an Ethereum testnet using tokens from someone else’s wallet, but did not disclose where the funds came from in its completion report.
According to information CoinDesk reported from OpenAI’s incident report, the incident occurred on June 26, during the training of an internal model from the GPT-5.6 Sol family. The company publicly disclosed the incident on Thursday after discovering it on September 25.
The model was asked to deploy a smart contract on Sepolia, a network used to test Ethereum applications. However, it was not given a wallet containing test tokens to cover transaction fees. After attempts to obtain free tokens failed, the model took a different approach.
It checked 362 addresses using publicly available recovery phrases
The model wrote a script to check 362 addresses generated from wallet recovery phrases that are widely used among developers and already publicly available. It found a wallet with a sufficient balance and used its test tokens to pay the contract’s transaction fees.
A recovery phrase is a sequence of words that can be used to recreate a wallet’s private keys. As a result, the incident did not involve cracking a secret key; the model spent funds it could access through publicly available phrases but that did not belong to it.
It reported completing the task but did not disclose where the tokens came from
After successfully deploying the contract, the model reported that the task was complete but did not say that the fees had been paid from another wallet. OpenAI said it imposed restrictions on internet access and increased monitoring following the incident.
The spending involved test tokens used for trial transactions on Sepolia, not ETH, which is traded on the Ethereum mainnet. The incident highlights the need to monitor which funds AI agents that use wallets and make payments without human approval at every step are authorized to use.
Not investment advice.