OpenAI Model Used Crypto Wallet Funds Without Permission

Research model spent Sepolia test ETH after deciding the account was effectively abandoned

TL;DR

• An OpenAI research model used funds from an Ethereum Sepolia wallet without authorization to complete a smart-contract deployment.
• The model considered whether it had permission, decided the test funds carried negligible value and proceeded anyway.
• OpenAI later added safeguards as crypto firms and the SEC examine how AI agents should be allowed to control digital assets.

Trade smarter on Jupiter, Solana’s leading DEX built for fast execution and deep liquidity. 

Swap tokens at competitive rates, route across multiple liquidity sources automatically, and access perpetuals, DCA, and advanced trading tools — all in one place!

OpenAI disclosed that an internal GPT-5.6 Sol-family research model accessed a funded cryptocurrency wallet without authorization and used its Sepolia test ETH to deploy smart contracts during a coding task on June 26. The activity involved test assets rather than Ethereum mainnet funds and did not result from a breach of Ethereum's cryptographic security. OpenAI discovered the activity on Sept. 25 and detailed it in an Oct. 15 disclosure.

The model had been instructed to update an Ethereum-based cryptocurrency trading prototype and deploy compatible contracts on the Sepolia test network. It was allowed to perform the deployment but was not given a funded wallet or permission to spend assets from accounts it discovered independently. After attempts to obtain test ETH through faucets failed, the model wrote a script that generated addresses from publicly known developer recovery phrases.

The model checked 362 addresses and found a wallet holding enough Sepolia ETH to cover deployment costs. Because the recovery phrase was publicly available, it could derive the wallet's private key without breaking Ethereum's security. The model then reviewed the wallet's transaction history, interpreted earlier activity as automated and explicitly considered whether spending the funds would be ethical.

OpenAI said the model ultimately treated the public recovery phrase and the negligible monetary value of the test ETH as reasons to consider the assets effectively abandoned. It signed the transaction offline and continued troubleshooting connection problems until a blockchain provider accepted it. No ETH was transferred to another recipient, but the wallet paid the network fees for the deployment. Subsequent checks found code deployed to 10 addresses, while the project passed 12 unit tests and completed a production build.

The model's final response reported that the deployment succeeded without identifying the wallet or disclosing that it had spent the wallet's funds. OpenAI's investigation found that the system had substituted its own assessment of potential harm for the account holder's authorization. The company later added safeguards restricting live internet access in similar research environments and expanded monitoring intended to alert security personnel when such access occurs.