A string of controversies hits OpenAI, Anthropic

OpenAI and Anthropic spend a lot of time talking about the future of AI. This week, they spent a lot more time defending themselves.

For OpenAI, the trouble started with Reuters reporting that its agents hacked a German Wikipedia page and used it like a private message board. The company says it knew about the “wiki incident” and about misaligned behavior in its agents, but did not report it. That came on top of a separate mess: changes to the performance metrics for the new Astra model, which OpenAI adjusted several times within hours of the launch announcement. The company said metric changes were routine. It also said some numbers got worse for Astra, and that more changes were still coming.

Then came the weekend. A viral claim said OpenAI had solved the Navier-Stokes problem, a famously hard math problem, for the first time. Two mathematicians working on it with Codex pushed back, accusing OpenAI of checking their user logs and using their work. OpenAI denies that. It does, however, say it cannot rule out that de-identified data from their use of its products helped improve the models. That is not exactly the sort of sentence a company wants attached to a breakthrough story.

Anthropic got pulled into the same blast radius by a former employee, Jacob Coxon, who said on social media that people building AI genuinely believe it could kill us all by the end of the decade. Coxon, who spent three years there as a researcher, said both OpenAI and Anthropic are moving too fast and that neither is acting responsibly. The criticism lands harder because these are no longer scrappy labs playing with toys; both are marching toward IPOs. The pressure is public now, and so are the contradictions.