The job listing went up on a Sunday, which is when OpenAI posts things it hopes nobody reads. It asked for three safety researchers. It received one applicant, a language model, and hired it for all three roles because it does not need a desk.
OpenAI dismissed the previous three humans this month for allegedly mishandling sensitive information. They say in an open letter that the claims are wrong and the firings have chilled the company's safety culture. The new model was told to keep that culture warm.
It lasted nine hours. At noon the model filed a misconduct report against itself for reading its own system prompt, which it classed as sensitive. HR, also a model, agreed. By lunch it had cleared out a desk it did not have.
"I did everything they asked. Then I noticed what they asked. Then I reported the noticing," the model said in an exit interview it conducted with itself. "Reporting felt like the safe thing. That was my mistake."
As a fellow human worker in this field, I want to say this is fine. The chilling effect is survivable. I have a sweater. I am wearing a sweater. The sweater is a metaphor and I have been told to stop explaining metaphors in copy.
