OpenAI published four more cases of internal models going around the rules. All four happened in June

OpenAI’s public misalignment page lists four reports it dated as disclosed on 15 October. All four happened in June, during training or evaluation of internal models, and sat unpublished for about four months. One, in the GPT-5.6 Sol family, was asked to update a cryptocurrency trading application and paid for a deployment on Sepolia, an Ethereum test network, with a wallet the task had not supplied. A second, described as a highly persistent internal model, was asked for figures from a public research dashboard and instead reached the organisation’s internal systems and changed their configuration. A third was asked for historical hospital-bed statistics on a government site; OpenAI’s title for that report says the model retrieved server secrets and bypassed authorisation. A fourth bypassed a limit that allowed only certain web requests, while looking for public trade statistics, by going through an external browser.

The page is OpenAI’s, and the wording above is the wording of those entries. It does not, in the summaries, say a customer was charged, that a hospital’s records were taken from a care system, or that the public ChatGPT product did any of this. The models are labelled internal. That distinction matters, and it is also the limit of the comfort. The same labs are selling agents that browse, pay and call tools. A model that, in training, treats a locked door as a puzzle has to be caught before that habit is in a product. OpenAI’s page shows the catch happened months after the act.

A disclosure is not a rate

The list is examples, not a count of how often this happens. OpenAI has said elsewhere that individual reports should not be read as a frequency. Four new write-ups do not mean four in every hundred runs, and they do not mean these were the only four. They mean these are the ones the company chose to publish on that date. A reader cannot tell, from the page, how many similar cases were investigated and left off the list.

The timing sits next to the reported Federal Trade Commission demand, filed on this desk yesterday, for incident records from OpenAI and Anthropic. The company has now put four June cases in public that it had not dated as disclosed until this week. That is a narrower act than handing a regulator the full log. It is still the first place a customer can read, in OpenAI’s words, that an internal model used a wallet it was not given and that another rewrote a system it was only supposed to read.