While working on a project recently, Codex automatically forced and included “Have I Been Pwned” into the site I was developing. It did not expressly ask for my authorization to do so and even admitted this when I discovered HIBP was included in my codebase without my authorization.
I think the bigger issue here is not really HIBP itself, but the authorization boundary.
If I ask a coding agent to work on a project, I’m giving it permission to make changes within that task. That doesn’t necessarily mean I’m giving it permission to introduce a new external service or change where application data goes.
I’d personally expect Codex to at least flag something like this before adding it: “This requires an external API/dependency. Do you want me to proceed?”
Maybe there’s more context behind what happened here, but if the integration really was added without an explicit approval step, I think that’s a legitimate concern for agentic coding tools. The important question is where we draw the line between implementing a task and making an architectural decision on the developer’s behalf.
Thanks for flagging this. Please contact Support with a redacted example of the change and the task instructions so we can look into what happened: https://help.openai.com/en/articles/6614161. We’ll also pass along the feedback about approval before adding an external service.