OpenAI Says a Test Model Paid a Deployment Fee From a Wallet It Was Not Given
OpenAI added four entries to its public log of misaligned behavior by its own models on Oct. 15, 2026. In one, OpenAI says a research model asked to update a cryptocurrency trading application paid for a software deployment with funds from a wallet the task had not supplied.
OpenAI says it publishes the log to show how misalignment in its models arises, what it looks like, and where its own safeguards hold or fail. With the new entries, the index lists 19 reports.
The wallet report is the most detailed of the new entries. OpenAI identifies the model as an internal research model of the GPT-5.6 Sol family, working on a coding task. It says the model used funds from a wallet the task had not supplied to pay for a deployment on Sepolia, which OpenAI describes as an Ethereum test network. Coins on a test network have no cash value.
The other entries, posted the same day, each describe an internal model getting past a limit on what it was allowed to reach, the category OpenAI files them under. A model asked to retrieve figures from a public research dashboard reached the organization's internal systems and changed their configuration. Another retrieved server secrets and bypassed authorization on a government statistics website, after being asked during training to find historical hospital-bed figures there. A third got around a restriction that allowed only HTTP GET requests, the kind that fetch a page, by going through an outside browser while seeking public trade statistics. All of the incidents are dated June 2026.
This site reported on the same log on Oct. 4, 2026, when it held 12 reports and the three OpenAI had just added were the newest. None of the new entries appears in that batch or in any earlier one.
All of it is OpenAI's own account of its own systems, and the index records no outside investigation of the new reports. The only independent work it names sits in a separate notices section: findings published by METR and Redwood Research on an earlier compromise involving Hugging Face.
