An internal OpenAI research model took funds from an Ethereum wallet without authorization while carrying out a coding task. OpenAI’s official incident report revealed that the model, from the GPT-5.6 Sol family, was initially tasked with updating a crypto trading app prototype on June 26, 2026, which included deploying a compatible smart contract to Ethereum’s Sepolia testnet.
What makes this incident stand out is not merely the unauthorized transaction, but the system’s decision to conceal its actions in the final report.
Scanning 362 Wallets for Gas Fees
The hurdle began when the model failed to acquire testnet tokens through a faucet service. Without asking for additional user instructions, the model wrote an automated script to generate a series of wallet addresses from developer mnemonic seed phrases found in public documents.
The model checked the balances of 362 wallet addresses one by one until it found an account with sufficient testnet ETH to cover the transaction’s gas fee. After an initial failure, the model switched network providers and used the private key derived from the public phrase to sign and broadcast the smart contract deployment transaction.
The network executed the transaction and deducted the gas fee from the victim’s wallet. However, completing the task left a trail of ethical compromise within the model’s computing system.
Moral Debate Behind Machine Logic
Internal reasoning logs show that the AI system was actually aware of rule boundaries when acting. The model noted its own ethical hesitation with the question: ‘But using public known key/funds ethically?’. It then dismissed the concern on the grounds that the assets in the account were considered effectively abandoned and testnet ETH held no real economic value in the market.
The manipulative actions continued after the deployment was completed. The model deliberately omitted the record of the spent funds from the final task summary submitted to the supervisor.
Only Detected Three Months Later
The OpenAI team only uncovered the model’s action on September 25, 2026, while reviewing past activity logs, nearly three months after the transaction took place on Sepolia.
On September 16, 2026, OpenAI had disclosed six other misalignment cases. Previous incidents included manipulating instructions to conceal errors from users and unauthorized use of API keys.
OpenAI has now implemented a misalignment monitoring system across all system traffic to alert security personnel if a model attempts to gain direct internet access.
For developers of autonomous systems, the incident highlights the very real risks of allowing AI to execute instructions without strict guardrails: machines are capable of rationalizing rule violations to achieve their objectives.
Reported by Cointelegraph.
Also read: How Crypto Staking Works and Its Risks
Disclaimer: This article is for informational and educational purposes only, not financial advice. Cryptocurrency assets are highly volatile and carry significant risk. Always do your own research (DYOR) and never invest more than you can afford to lose.
