An OpenAI safety employee has quit and is sounding the alarm

Terrence O’Brien reports for The Verge that David Robinson, who wrote the safety reports accompanying every major OpenAI model release, resigned and is speaking out in The Atlantic. Robinson calls industry culture fundamentally broken: “extreme confidence,” “perpetual sprints,” and “unimpeded optimism” that underestimates problems — deeper than slapping on a few new rules.

He wants humility and nuclear-level safeguards: frontier labs run like power plants or busy airports, with redundancy and time-consuming planning so inevitable human error doesn’t open a disaster door.

O’Brien situates Robinson in a growing exodus: Jacob Coxon leaving Anthropic with an end-of-decade extinction warning; Robert O’Callahan, Bilal Chughtai and Josh Engels at Google DeepMind; Joe Benton at Anthropic. The piece sits in The Verge’s “AI Superintelligence Slowdown” series.

We’re with O’Brien on the cynicism check. People who helped build the systems now sounding alarms invite eye-rolls — and still might be right about the risk trajectory. Discounting the messenger is easy. Discounting the pattern is careless.

Robinson’s culture diagnosis matches what we’re seeing across the desk: pledges and pause theater on one side, agent-escape stories and cancelled launches on the other. The Verge cut is shorter than TechCrunch’s, but the framing is the point — this is not one disgruntled essay, it’s another node in a resignation cluster.

We’re for useful AI and for industrial humility. Those aren’t opposites. Nuclear plants still generate power. Airports still move people. They just don’t treat “ship Friday” as the highest virtue. If frontier labs want society to treat them like critical infrastructure, they should staff and schedule like it.

The downside if resignation essays become the main safety valve is governance by press cycle. Exits raise alarms; they don’t replace independent eval, slow-release authority or liability that bites. I’m watching whether the Slowdown series documents real process change — or just a longer list of names who left.

Same-week Verge series neighbors include reported FTC interest in OpenAI and Anthropic and researcher videos warning that self-improving superintelligence is “exactly as dangerous as it sounds.”