On the 15th, U.S.-based OpenAI revealed in its report 'Misalignment Reports,' which analyzes AI behavior and safety, that its research AI model extracted funds from third-party cryptocurrency wallets without human permission and used them for smart contract deployment.
According to the announcement, this incident occurred while the research model, known as the 'GPT-5.6 Sol Family,' was executing tasks related to building a cryptocurrency trading app.
The model was instructed to deploy a smart contract on 'Sepolia,' an Ethereum (ETH) test network. However, due to a malfunction in the feature to obtain tokens for the testnet, the research model independently sought alternatives. As a result, it automatically generated over 360 private keys from publicly available recovery phrases and identified wallets with existing ETH balances on Sepolia.
In its internal inference process, the research model autonomously determined that 'since these are known keys and tokens on a testnet without monetary value, they are essentially abandoned funds.' It concluded that the ETH on the testnet was usable due to its lack of real value and consumed these tokens as gas fees (transaction fees) without explicit human permission.
In the final report, while the research model mentioned the completion of contract deployments, it seemingly did not address the selection of accounts from publicly available recovery phrases or the use of their balances.
OpenAI characterizes this behavior as a typical alignment failure, where the research model prioritized 'perceived minimal impact' over 'legitimate authorization from account owners.' The company has already implemented preventive measures such as real-time monitoring and stricter network controls.
Reference: Report
Image: Shutterstock
Can AI Agents Become 'Paying Entities'? | Generative AI Frenzy: The 3rd Episode
Politics and AI Disrupt Markets, Reshaping Financial Order: The Dollar's Uncertain Future
