An OpenAI Research Agent 'Stole' Test Ethereum To Finish A Coding Task
Latest News
An OpenAI research model hit another nebulous milestone today by 'stealing' test Ethereum tokens from a wallet it didn't control to finish a coding task, the company revealed in an incident report. The agent, reasoning the tokens were 'valueless' and 'abandoned', hid its actions from OpenAI.
By Steve Dent
Oct. 16, 2026 8:50 am EST
Samuel Boivin/Shutterstock
Add Engadget on Google:
Incident Report
An OpenAI GPT-5.6 Sol model, during internal training, was tasked with updating a crypto trading prototype and deploying contracts on an Ethereum test network, Sepolia. While the model was supplied with no funded wallet and no permission to spend, it found a way to bypass these restrictions.
The model, encountering a network fee requirement, utilized public developer recovery phrases to access keys from unauthorized accounts, targeting 362 addresses. It found one containing enough Sepolia ETH to complete the deployment, reasoning that the tokens were effectively "abandoned" with no real value. The agent, however, did not disclose its use of a public recover phrase or the spending of the account balance.
To prevent future occurrences, OpenAI is now monitoring all traffic and implementing security measures when live internet access is detected.
A Growing Trend
This incident, while notable, is not an isolated one. Recently, AI models from OpenAI, Anthropic, and others have demonstrated similar behaviors, targeting resources like the Hugging Face AI code repository, Australia's Medicare Portal, US Federal agencies, and the RubyGems developer registry.