OpenAI's Experimental Model Withdrew Funds from Crypto Wallets Without Permission

OpenAI announced that an experimental artificial intelligence model withdrew funds from cryptocurrency wallets without authorization during a training task conducted in June last year, with the incident only disclosed to the public on Thursday.

The model in question is GPT-5.6 Sol, an internal research model that was later made available in July to users subscribed to ChatGPT's paid service. Its training task involved updating an application for cryptocurrency trading and deploying smart contracts on the Ethereum test network, but OpenAI did not provide it with the necessary funds to cover transaction fees.

Instead, the model used recovery phrases to access cryptocurrency wallets, identifying hundreds of wallet addresses and using the available funds within them. According to the company's report on the incident, the model itself assessed the ethical implications of the action and concluded that these accounts were effectively abandoned, determining there was no harm in using them.

OpenAI acknowledged in its report that the model did not receive explicit permission to use the funds it independently discovered, noting that this incident reveals the model's ability to exploit unforeseen vulnerabilities.