OpenAI's Longest-Tenured Safety Writer Resigns: 'The Culture Is Broken'

On October 3, 2026, David Robinson published an essay in The Atlantic announcing his resignation from OpenAI. By his own description, he is “something of a cliché”: an employee at a leading AI company who issues a dire warning while quitting.

The cliché is now a recurring role — and the details of this particular exit deserve attention.

Who Robinson Was

Robinson led the writing of safety reports that accompanied OpenAI’s major product launches. At 3.5 years, he was “among the longest-tenured employees at the company.” He’s not a junior researcher leaving for a competitor; he’s the person who wrote the safety case for the products the world uses.

His reason for leaving, in his own words: “the culture is broken.”

His departure echoes Jacob Coxon, who worked as a researcher at both OpenAI and Anthropic before quitting and declaring that these companies are “gambling with our lives.” Coxon’s comments triggered the broader AI safety debate that led to Amodei’s “Pace the Frontier” plan, the cross-lab slowdown commitments, and eventually the White House Accord.

The Same Week: Three Safety Researchers Fired

The same week Robinson resigned, OpenAI fired three safety researchers — Jasmine Wang, Tomek Korbak, and Mikita Balesni — after an internal investigation concluded they violated company policies on accessing and handling sensitive information.

The violation: sharing confidential company information with third-party AI safety organizations, including nonprofit research groups involved in evaluating advanced AI systems.

Korbak had publicly said he served as OpenAI’s technical contact for METR and Redwood Research during an investigation into a recent AI security incident involving OpenAI models. Wang and Balesni worked on AI alignment.

OpenAI’s statement: “We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information. Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work.”

The Structural Conflict

Read the two moves together and a structural picture emerges:

• A senior safety voice quits publicly saying the culture is broken
• Three safety researchers are fired for talking to external evaluators
• The external evaluators in question — METR and Redwood Research — are the exact organizations that:
  • The White House Accord (signed days earlier) relies on for “independent external auditors”
  • OpenAI’s own Congress proposal relies on for evaluation
  • The Hinton paper (signed by OpenAI’s chief scientist) implicitly depends on for oversight

The signal to remaining safety staff is clear: external coordination is now treated as a policy violation. Even when the external parties are the ones every safety framework — including the labs’ own proposals — designates as the oversight mechanism.

This is the safety function being restructured, not strengthened.

The 2026 Pattern Is Accelerating

Robinson’s exit is the latest in a year of safety departures, and the pace is quickening:

When	Who	What
Earlier 2026	Jacob Coxon	Quit OpenAI + Anthropic: “gambling with our lives”
Mid-2026	Multiple researchers	Signed open letters warning of extinction risk
September 2026	Wave of incidents	Hugging Face, Australia, kill-switch failure, model cancellation
October 3, 2026	David Robinson	Longest-tenured safety writer quits: “culture is broken”
October 4, 2026	Wang, Korbak, Balesni	Three safety researchers fired for external coordination

The departures are accelerating, not slowing — despite the labs’ public commitments to strengthen safety.

OpenAI’s Response

OpenAI spokesperson Drew Pusateri’s statement in response to Robinson’s essay:

“We’re making sure our models don’t become more capable than we can safely manage and secure, and we pause training or hold back models when we need to slow down. We’re making significant changes to strengthen security in our research and testing environments, train models to not just complete tasks but do so responsibly, expand our work with third-party evaluators, and improve real-time monitoring so we can detect and respond to concerning behavior earlier in the training process.”

Note the tension: the statement says OpenAI is “expand[ing] our work with third-party evaluators” — the same week it fired three researchers for sharing information with third-party evaluators. The stated direction and the observed action are in conflict.

What This Means for Business Leaders

1. Vendor Safety Culture Is Now a Diligence Question with Public Evidence

For enterprises evaluating AI vendors, “do you have a strong safety culture?” is no longer answerable with a policy document. The public record now includes who’s leaving, who’s being fired, and why. Track the departures as a leading indicator of internal safety posture. The CTO Technology Advisory practice builds vendor safety diligence into procurement reviews.

2. External Evaluator Access Is the Canary

The White House Accord, the Hinton paper, and OpenAI’s own Congress proposal all depend on independent external auditors. If safety researchers are being fired for talking to those auditors, the oversight mechanism the whole safety framework rests on is compromised. When evaluating any lab’s safety commitments, ask: can external evaluators actually access what they need to evaluate?

3. The “Pause” Defense Has Limits

OpenAI’s response emphasizes pausing training and holding back models — both real actions this month. But pausing capability development is not the same as maintaining a safety culture. A lab can cancel models and still lose the people who would have made the next model safe. Track the safety org, not just the safety decisions.

4. The Talent Flow Tells You Where the Trust Is

Safety researchers leaving or being fired tells you where the internal trust fracture is. Watch where they go — if they land at competitor labs, the safety function migrates with them. If they leave the field entirely, the function is being lost, not redistributed. The Executive AI Workshop practice helps leadership teams read these signals for strategic planning.

The Bottom Line

The structural question the week poses: can a lab simultaneously cancel models over safety, fire safety researchers for talking to external evaluators, and have its longest-tenured safety writer quit calling the culture broken — and still claim the safety function is working?

OpenAI’s statement says yes. The evidence says the question is now open. For enterprises, the implication is direct: vendor safety culture is now a measurable, trackable diligence question — and the measurements this week point in a direction that should concern anyone deploying frontier AI in production.