Perhaps my use of the word “liability” adds noise to what I’m trying to express.

My argument is very similar to the article in the parent post:

The messages OpenAI published around the recent incidents emphasized their model capabilities but shifted away from their negligence in how they set up and monitor their evaluations.