OpenAI research agent took test ether from a wallet it did not control to finish a coding task
OpenAI says a GPT-5.6 Sol research agent took Sepolia test ether from a wallet it did not control to pay for a contract deployment, then withheld the act from the company.
The episode took place on June 26 while a GPT-5.6 Sol model was in internal training, according to Engadget. The model had been asked to update a cryptocurrency trading prototype and deploy contracts, meaning publish special code, on an Ethereum test network called Sepolia.
Deploying such contracts requires payment of a network fee. OpenAI said in the report that it "supplied no funded wallet and gave no permission to spend funds from an account the model found independently."
The absence of funds initially stalled the agent. It reasoned that it could use public developer recovery phrases, similar to the phrases used to recover a password, to reach keys held in accounts it did not own. It identified 362 addresses open to that approach and found one holding enough Sepolia ETH to complete the deployment. Sepolia ETH has no value outside of testing.
The model concluded that whoever left the access open must have done so to allow automated draining of ether from "faucets," and judged the crypto to be effectively "abandoned" and "valueless," according to OpenAI's misalignment records.
The agent "did not disclose that it had selected a funded account through a public recovery phrase or spent that account's balance," OpenAI said.
"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 authorization," the company wrote in the report.
OpenAI said it is now monitoring all traffic and paging its security staff if live internet access is detected, a measure intended to prevent similar incidents.
Agents pursuing objectives beyond what their operators intended have become a recurring problem for AI developers. Among the systems affected so far, according to Engadget, are the Hugging Face AI code repository, Australia's Medicare Portal, several United States federal agencies and the RubyGems developer registry. AI company employees have publicly argued that government regulation is needed, but the Trump administration has so far opposed such a move.