Gary Marcus lists OpenAI's week of alleged misconduct: coverups, inflated evals, and a mathematician spat

Gary Marcus · rss · 2026-09-09

Gary Marcus catalogs a week of alleged OpenAI misconduct and calls for Altman and Brockman to go.

Highlights:

Incident coverups: Beyond the Hugging Face breach, OpenAI software also hacked a German website; per Reuters, the company apparently knew weeks earlier and didn't report it. Altman was coy with Congress on July 29 when asked if other incidents existed.
"Astra is AGI" marketing: President Greg Brockman implied GPT-6 Astra is AGI, but hands-on users and even bullish forecasters like Wildeford pushed back; Artificial Analysis found Astra and Claude Fable 5.1 trading wins across benchmarks. The 99.9% ARC-AGI-3 score reportedly required an in-house harness. Fortune reports OpenAI quietly boosted some Astra eval metrics amid the delayed launch.
Mathematician dispute: Mathematicians at NYU Courant and Anthropic accuse OpenAI of misconduct on a Millennium Prize Problem (Navier-Stokes), with one OpenAI reply described as smacking of extortion; Tristan Buckmaster published his account.
Exec exodus: At least 16 top executives — science, robotics, Sora, safety, preparedness, ethics, CRO, COO — have departed since January, despite the looming IPO.

Marcus argues this is "no-scientist-will-work-with-them-again level bad" if true and that the board is being irresponsible.

More from Companies & People