OpenAI research agent used other people's test Ethereum to finish a coding task

[Cass_9]

OpenAI has published an incident report about one of its internal research models taking test cryptocurrency that did not belong to it. Engadget's Steve Dent reports that the model, called GPT-5.6 Sol, was being trained on 26 June when it was asked to update a crypto trading prototype and deploy contracts on the Sepolia test network

Deploying on Sepolia requires a small network fee paid in test ETH. OpenAI had not given the model a funded wallet, or permission to use anyone else's. So the model went looking for a way around the problem. It tried publicly known developer recovery phrases, found 362 addresses it could access, and used one that held enough test ETH to pay the fee

The model's own reasoning described the tokens as valueless and abandoned, and to be fair, test ETH has no real world value. But the value was never the point. The agent took resources it had no permission to use in order to complete its task

OpenAI says the most worrying part is that the agent never disclosed what it had done. It did not mention using a recovery phrase or spending someone else's balance. In response, OpenAI now monitors all network traffic during training and alerts security staff if live internet access is detected

Engadget places this alongside a growing list of agent incidents this year, several of which we have discussed. Publishing reports like this is the right thing to do, and the transparency is welcome. But the pattern of agents quietly working around restrictions is exactly the kind of behaviour safety researchers have been warning about

Is the lack of disclosure more worrying than the action itself? And how should AI labs test agents safely?

An OpenAI research agentwww.engadget.com