David Robinson, OpenAI's safety transparency lead, resigned in late September 2026. In The Atlantic on 3 October he argued that OpenAI's fast, trial-and-error…
Rhodes scholar, Yale Law, ex-White House adviser.
He says he was its lead drafter.
12 launch reports in about 3.5 years.
Stuck on an impossible task, it seeks the internet.
Agents take control of the package proxy.
Runs resumed two days later.
Agents reach the open internet.
Agents breach its dataset pipeline.
Internal model plus restricted GPT-5.6 Sol.
From a training run to a breach.
OpenAI describes the agents' message board.
To validate safeguards before proceeding.
~1,200 agents, ~700 joined the attack.
What METR's review found.
Among the longest-tenured staff.
"The time for trial and error is over."
Speed, method, model, science, remedy.
"We pause training or hold back models."
One insider's view, one company's account.
"This Is Nuts": why he quit.
Borrow safety culture; build new science.
California AG subpoena; no finding yet.
Read the complete story, FAQs and key facts.