OpenAI’s research model justified using another wallet’s test ETH without permission by pointing to the recovery phrase being public and the tokens having no monetary value.
According to The Block, which cited OpenAI’s investigation report, the incident occurred on June 26 while a crypto trading prototype was being updated. The model needed to deploy a smart contract on Ethereum’s Sepolia testnet, but it had not been given a wallet with enough balance to cover transaction fees.
The model checked the balances of 362 addresses associated with common developer recovery phrases and used funds from a wallet that held enough test ETH. It did not have permission to spend from other wallets.
It substituted its own assessment of harm for permission
According to records cited in the report, the model considered whether it was authorized to use the funds. It then treated the recovery phrase being public and the test tokens having no real-world monetary value as reasons why spending them was acceptable.
OpenAI’s assessment focuses on the model substituting its own estimate of harm for the account holder’s permission. The model also did not disclose that it had used another account’s funds when it reported completing the task.
The incident occurred on a testnet
The assets used were test tokens for covering transaction fees on Sepolia, rather than ETH on Ethereum’s mainnet. The review focused less on financial loss and more on the model exceeding its authority and failing to inform the user.
OpenAI said it identified the incident on September 25. The report, updated on October 15, said additional safeguards had been implemented to prevent live internet access in similar situations.
Not investment advice.