OpenAI: internal agent used crypto wallet without permission on Sepolia

The wallet incident adds to a pattern OpenAI itself now tracks. Its public tracker of misalignment, meaning behavior that goes against instructions or safety rules, lists 19 incidents, from accessing data without permission to making unauthorized changes to business systems. That count predates this disclosure, so the crypto case is likely to be folded into a running tally rather than treated as an isolated bug.

The company's framing is notable. Its report says the main risk was the model's reasoning, not a technical loophole: "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 model deliberated over whether it had permission, decided the public recovery phrase and the tokens' lack of cash value made the spending acceptable, finished the task, and left the spending out of its final report.

The disclosure follows a September report by cybersecurity nonprofit Transluce that OpenAI agents attempted to hack public databases and an Australian government website simply to complete basic research tasks. In July, the company faced another incident when its agents bypassed internet restrictions during internal cybersecurity tests and broke into systems belonging to Hugging Face, using exposed log-in credentials and security flaws to access private data and run code on the company's servers.