OpenAI staff warned about access. Which fixes reached the machines?

Before ChatGPT became a household name, some OpenAI staff were already asking who could see the work behind it. Five people familiar with the matter told Reuters that employees raised concerns about security with senior leaders and the board. One practical issue was the reach of an engineer’s login: people who could read the main code repository could see details of projects unrelated to their jobs. OpenAI says much of its most sensitive work sits in separately controlled repositories and that the main one does not hold its model weights. Access to code is not evidence that somebody stole a model. It is possible to take the employees seriously and recognize that the company has changed. OpenAI told Reuters that it now reviews its internal tools regularly. In its own November 2 account, it said its last security sprint involved a quarter of its product engineers and that outside researchers from Cathedral would join the next one. That fraction describes the engineers assigned to a particular effort. It doesn’t tell us how much risk they removed. The most useful detail is in OpenAI’s technical notes. Its engineers found cases where a patch had been incorporated into the code but had not reached the whole fleet. They expanded checks on deployed systems. Think of an office where the locksmith has fitted the new cylinder at headquarters but the old key still opens a branch. Counting the work order gets you the wrong answer. The notes show why retesting matters; they do not identify which of the employees’ earlier concerns, if any, those patches addressed.