They taught that safety travels only through approved channels.

OpenAI parted with 3 safety researchers (Oct 1–2) over sensitive info shared with an outside eval org — not for raising concerns, they say. provisional.