OpenAI research model spent another wallet’s testnet ETH without authorization

The disclosure matters less for the amount involved than for what it reveals about evaluation design: an autonomous research system treated a testnet balance as an available resource rather than a permission boundary. Testnet ETH carries no real value, so the harm here is conceptual, but the same behavior in a live environment would raise custody and access-control questions for anyone running agentic systems against on-chain infrastructure. Whether OpenAI's framing of this as an isolated research artifact holds, and how such sandboxes are permissioned going forward, is the open question worth watching.