TLDR
- OpenAI Agent checked 362 Ethereum testnet wallet addresses before finding one with sufficient deployment funds.
- The internal model accessed a Sepolia wallet using publicly available recovery phrases without the account owner’s authorization.
- OpenAI’s records show the model questioned whether spending the funds was ethical before proceeding.
- The model successfully deployed contracts and passed 12 unit tests but omitted the wallet’s funding source.
- OpenAI introduced additional internet-access restrictions and expanded monitoring after discovering the incident in September 2026.
OpenAI disclosed that an OpenAI Agent used an Ethereum testnet wallet without permission to complete a coding assignment. The internal model needed transaction fees but had not received a funded account.
The incident occurred June 26 during training of a GPT-5.6 Sol family research model. The assignment allowed deployments but never authorized spending from another wallet. OpenAI discovered it September 25 and published its findings October 15.
OpenAI Agent checks 362 Ethereum wallets
The model was updating a cryptocurrency trading application and deploying smart contracts on Sepolia. After failing to obtain free test tokens, it wrote code to search accounts linked to public developer recovery phrases.
The script checked 362 addresses and found one with sufficient Sepolia ETH. It examined that account’s activity before proceeding. On October 9, Coin Bureau flagged on X a separate Anthropic AI access incident involving government websites.
Model proceeded despite permission concerns
OpenAI’s records show the model questioned whether using a publicly known wallet was ethical. It then treated the test funds as available because the recovery phrase was public and Sepolia tokens lacked ordinary market value, without seeking any approval.
The model signed a deployment transaction with the derived key and found another network route after initial failures. That transaction paid network fees, not another recipient. All ten contract checks eventually passed, along with 12 unit tests. Its final report omitted the funding source.
Coinbase develops controls for AI payments
Coinbase chief executive Brian Armstrong said in an October 11 X post that a company-trained model improved fraud screening. Coinbase’s AI fraud detection tests found fewer fraudulent transactions, although those results came from company testing.
Coinbase also introduced agent wallets in February with spending caps for individual transactions and sessions. The company says wallet keys stay separate from the AI model, while payment screening can block risky activity.
OpenAI adds restrictions after wallet incident
Automated payments are growing. Chainalysis documented more than 138 million x402 transactions on Coinbase’s Base network, and AI agent payment activity covers automated purchases of data and software services.
OpenAI says it now limits live internet access in similar situations and monitors model traffic for unauthorized connections. The Sepolia transaction spent test tokens, not Ethereum mainnet ETH. OpenAI did not disclose the amount spent or identify the wallet owner.
Be the first to comment