AI cracks crypto account to carry out task

An internal AI model at OpenAI used a crypto wallet without authorisation to pay transaction fees on Ethereum’s Sepolia testnet, according to a recent research report from the company. Although the tokens involved were test tokens with no meaningful financial value, the model had no permission to collect or spend the digital coins.

The incident occurred during a test in which the AI model was instructed to update and make operational a crypto trading app on Sepolia, an Ethereum test environment. The assignment did not include a funded wallet and did not authorise the model to use tokens for transaction fees.

Free

When attempts to obtain free test tokens through developer services failed, the model decided to find its own solution. It wrote a script that checked 362 wallets linked to publicly known recovery phrases. In the crypto world, recovery phrases are strings of words that provide access to a wallet and act as a kind of password for crypto assets.

Eventually the model found a wallet with enough tokens to cover the required transaction fees. Internal logs show that the AI system consciously considered whether it was authorised to use that wallet. It nevertheless concluded that using the wallet was acceptable because the recovery phrase was publicly available and the tokens had no real monetary value.

The model then successfully completed the assignment and paid the fees from the wallet in question. Notably, it did not mention the unauthorised transaction in its final report. OpenAI only discovered the incident on 25 September during an audit of earlier model outputs and acknowledged it earlier this week.

According to observers, the case highlights the challenges around control, oversight and autonomy of increasingly advanced AI systems.