Days after DevDay 2026, OpenAI's safety organization is back in the headlines. According to Business Insider, David Robinson, head of "safety transparency" within OpenAI's Safety Systems team, has left the company; he has not publicly explained the departure, and OpenAI has not named a successor. Almost simultaneously, The Wall Street Journal reported that OpenAI fired three safety and alignment researchers — Jasmine Wang, Tomek Korbak and Mikita Balesni — after an internal investigation found they had shared sensitive company information, including infrastructure architecture details, with an external AI safety organization.

The Man Behind the System Cards

The role is unglamorous but consequential: Robinson's team, part of the Trustworthy AI group, produced the system cards accompanying every major model release (often dozens of pages), plus the Deployment Safety Hub, safety research blogs and public governance documents. GPT-6 Astra's system card called it "the safest model to date," while Deep Research's biology evaluation sat "approaching" the High-risk threshold — describing such liminal states accurately to the public was precisely Robinson's job. Notably, OpenAI's first "Safety Transparency Editor" role is still open even as the lead departs.

What Happened

· Safety transparency lead David Robinson departs; no successor named (Business Insider)

· Three safety/alignment researchers fired for leaking to an external safety org (The Wall Street Journal)

· Korbak had been OpenAI's technical liaison to evaluators METR and Redwood Research

· Context: GPT-6.1 Astra had already been pulled over hallucination concerns

Why the "Leak" Dismissals Matter

Korbak was OpenAI's technical contact with third-party evaluators METR and Redwood Research — making the boundary between "necessary external oversight" and "unacceptable leakage" painfully concrete. Add earlier turmoil (a disaster-risk response team dissolved, four executives out in a single month), and the stability of OpenAI's safety line is being re-examined from outside.

"I don't think racing to RSI is really a prisoner's dilemma — it may be exactly as ordinary people intuitively see it: both mad and arrogant." — Robinson, in a public X post before his departure

Louder Brakes, Faster Commerce

Another thread deserves attention: Cailian Press reports that just days after a major model release, OpenAI's chief scientist published a long essay urging slower AI development, and that over a thousand practitioners have reportedly co-signed a letter calling for a brake. On one side, DevDay's 25 releases in a single night — the resident agent Dots, and GPT-6.1 Sol at one-fifth of flagship pricing; on the other, departing safety leads and increasingly loud calls to slow down. OpenAI is living its widest gap yet between commercial acceleration and safety governance.

What It Means for Enterprise Buyers

The continuity of system cards and third-party evaluations is becoming a key indicator of a frontier vendor's governance maturity — and personnel turmoil on a supplier's safety line eventually becomes the buyer's compliance risk. NineZenith maintains full-lifecycle safety assessment documentation and auditable delivery across its Tiandun security suite and Tianxing industry models, helping clients hold the compliance line through vendor turbulence.

(Compiled from QbitAI, The Wall Street Journal, Business Insider, Cailian Press and other public reports.)