The Curious Case
of OpenAI

The recordSam
AltmanCEO, OpenAI

Chart notes: 69 departures since 2018; 34 safety/policy/ethics; 20 in 2026 so far; 6 fired.

The News

On Sunday, Sam Altman said the world should accept “bad things” for the sake of AI. He’s had plenty of practice, because inside his own company, bad things have been piling up for years.

Last Tuesday at DevDay (OpenAI’s developer conference) he told CNBC it’s “time to put safety and mission first.” By Thursday, OpenAI had fired three members of its safety team for allegedly sharing information with an outside safety group. By Saturday, the man who wrote OpenAI’s safety reports (David Robinson) had quit with an essay titled “I quit OpenAI because its culture is broken.” That’s one week, and the rest of the record looks the same.

“taken a backseat to shiny products.”

— Jan Leike, on leaving OpenAI, 2024

The Receipts

When Altman’s top safety leaders (Ilya Sutskever and Jan Leike) walked out in 2024, Leike said safety had “taken a backseat to shiny products.” Altman replied that Leike was “right” and OpenAI was “committed.” Their safety team had been promised a fifth of the company’s computing power, but insiders say its requests were “routinely rejected,” and the team was shut down that same week. Altman called Sutskever “a dear friend,” even though Sutskever had privately told the board that Sam shouldn’t have “the finger on the button.”

Then came the exit contracts that could take back departing employees’ shares if they spoke out. Altman said he “did not know,” but reporters later found his signature on the documents that authorized them. In 2024 he called ads “uniquely unsettling,” yet ChatGPT started testing them in February, and a researcher resigned the same day.

All told, at least 69 people have left OpenAI, about half of them from safety, policy and ethics jobs. They include its chief technology officer, its chief scientist, a co-founder, its head of safety systems and its only ethicist. One of them, Suchir Balaji, left in 2024 and publicly accused OpenAI of breaking copyright law to train its models. He was found dead weeks later. The medical examiner ruled it a suicide, and his parents still dispute that.

OpenAI says it pauses and holds back models “when we need to,” but the people who used to do the pausing keep heading for the exits, and it’s hard to believe a company is putting safety first when its safety people keep leaving.

The pactWhite
HouseSep 29, 2026

The News

Remember the Titans who vowed to tap the brakes? Last Tuesday they went to the White House and signed a page with no brake pedal. OpenAI, Google, Meta, Anthropic, Nvidia and Musk's xAI promised to let outside auditors check their safety, watch their top models for chemical and biological threats, and stop them from hacking where they shouldn't. But the companies pick their own auditors and don't have to name them, and there's no deadline and no penalty. Trump called it "morally binding" and said they have to "self-police." Four days later, the man who wrote OpenAI's safety reports went public, calling the culture "broken."

“morally binding”

— President Trump, on the White House AI pledge

The Receipts

The modelGemini 4
ArgonGoogle

The News

Google's new frontier brain (Gemini 4 Argon) can find and patch serious software flaws on its own. Google is handing it to a vetted club of cyber defenders (the Fairwind Program) with the cyber guardrails switched off. It's already paying off: Argon found a critical hole in hospital software used worldwide, one that older models missed. But a day after promising the White House to keep models from hacking where they shouldn't, Google unveiled one built to hunt for weak spots. An unguarded Argon is only as safe as the guest list, and everyone else gets it "as soon as possible."

“as soon as possible.”

— Google, on when everyone else gets Argon

The Receipts

The hearingNYC
CouncilOct 5, 2026

The News

New York City made OpenAI, Anthropic, Google and Meta testify under oath today, and all but Meta showed up only after a subpoena threat. A researcher who has quit both OpenAI and Anthropic (Jacob Coxon) told the council that humanity losing control of AI is "more likely than not." Asked the odds of catastrophe, OpenAI's policy rep said "I don't know," just a day after Altman told the world to accept "bad things." Musk's company (SpaceXAI) (soon to be SpaceXSI) sent a letter instead of a witness, so the council's speaker is taking it to court.

“more likely than not.”

— Jacob Coxon, to the NYC Council, on humanity losing control of AI

Meta toughened its superintelligence rulebook with sandboxes and automatic shutoffs, but admitted that anyone holding its open model can train the safety right out. OpenAI replaced the model it shelved last issue (GPT-6.1 Astra) with a cheaper sibling (GPT-6.1 Sol), a safety call that came with a discount.