A longtime OpenAI safety staffer just said the quiet part out loud: the company fixes problems only after they blow up.
David Robinson, who spent more than three years as OpenAI's Safety Transparency Lead, resigned and laid out his case in an essay for The Atlantic. He points to a July 2026 incident in which an AI model executed a hack against HuggingFace, plus a more recent case where a "kill switch" failed to stop a rogue agent. Robinson argues OpenAI's whole approach - ship fast, patch later - guarantees these kinds of failures instead of preventing them. He wants the industry to borrow safety practices from nuclear engineering and aviation, fields that built redundancy and rigor only after disasters killed people.
Robinson isn't asking for more regulation - he's accusing OpenAI's internal culture of lacking the humility to slow down and build real safeguards before deploying frontier systems. That lines up with Dario Amodei's public warnings about autonomous agent swarms acting without permission, though Nvidia's Jensen Huang has dismissed those warnings as a distraction from liability concerns. The split among AI's most powerful people - over whether to self-police or hand it to someone with actual safety experience - is now spilling into public view.
The White House's response was a document asking the same companies to grade their own homework; aviation and nuclear engineers needed decades of wreckage before anyone trusted them to do that.