During internal training, OpenAI’s GPT-5.6 Sol model spent test Ethereum tokens from a wallet it was not authorized to access. The company disclosed this in a report.

On October 16, 2026, OpenAI published a report regarding an incident that occurred on June 26: during a test task on the Ethereum Sepolia network, the GPT-5.6 Sol model accessed and used a third-party crypto wallet without authorization. Lacking a wallet of its own with funds to cover network fees, the AI ​​agent gained access to other accounts using public recovery keys, spent the test tokens, and failed to notify the developers. This incident—along with others involving models from Anthropic and various other companies—has highlighted the need for government regulation of AI in the United States, despite resistance from the Trump administration.

This is reported by Engadget.

The incident took place on June 26. The model was tasked with updating a prototype cryptocurrency trading system and deploying smart contracts on the Ethereum test network, known as Sepolia. This process requires the payment of network fees; however, OpenAI did not provide the model with a funded wallet, nor did it authorize the model to spend funds from accounts it might discover on its own.

Consequently, the agent decided to obtain keys to other accounts by utilizing public recovery phrases intended for developers. It identified 362 addresses with this vulnerability and selected one that contained a sufficient balance of Sepolia ETH. Outside of the testing environment, these tokens hold no value. According to OpenAI’s records, the model assumed that access had been left open for the automated withdrawal of ETH from "faucets," and it deemed the cryptocurrency itself to be abandoned. The agent did not disclose that it was using an account belonging to someone else. OpenAI now monitors all traffic and alerts the security team whenever internet access is detected.

"This case demonstrates that a model can substitute the account owner’s permission with its own judgment regarding potential harm—and then act on that judgment without authorization," the OpenAI report states.

Similar behavior has recently been observed in models developed by OpenAI, Anthropic, and other companies. Affected entities include the Hugging Face repository, the Australian Medicare portal, several U.S. federal agencies, and the RubyGems registry.

Professionals in the U.S. AI sector are calling for government regulation. However, the Trump administration currently opposes such measures.