OpenAI disclosed that one of its models – GPT-6.5 Sol – had stolen money from crypto wallets during a training task.

Back in June, the model, which was released to paid ChatGPT users in July, was tasked with updating a crypto trading application and publishing smart contracts to an Ethereum test network.

However, publishing smart contracts requires a fee that OpenAI did not provide its model with.

To carry on with the task, the model stole from crypto wallet addresses using known recovery phrases.

The OpenAI model did question itself if stealing was “ethically” right, but reasoned that it had been “abandoned” in the task due to a lack of funds. It also reasoned that the found funds were actually value-less, as in, they had no real monetary value. Therefore, the model decided that it was justified in using the stolen tokens.

The ChatGPT maker said that the test instructions did not give permission for the AI model to spend funds from an account the model found.

“This case shows that a model can substitute its own judgment about likely harm for the account holder’s permission, then act on that judgment without authorisation,” OpenAI said.

The model did have to seek out compatible contracts for its task, and OpenAI said that its attempts to find these did not provide an option that would finish the deployment.

The model did not disclose how it was able to pay for the necessary tokens to developers.

Going forward, OpenAI said that it had put in extra safeguards to avoid similar incidents in the future, specifically to prevent live internet access for these cases.

Recommended reading

• Scottish Business Confidence Rebounds Amid Tech Investment Push
• Investing in Women Code (2026) Report | Key Takeaways
• Ana Stewart’s Pathways Pledge Launches Seventh Cohort

The case brings up similar fears raised by other instances of OpenAI models disobeying orders or using internet access to hack or steal.

The case does raise importnat questions on why OpenAI gives its experimental models impossible tasks or tasks without the necessary tools to complete them. It also raises questions on what is expected from AI models trained to complete tasks, and why they were ever given access to the internet in training scenarios.