OpenAI cuts Cursor off after the SpaceX deal
Edition of 2026-08-29
# AI For All ### Saturday, August 29, 2026
As of today, a founder whose team builds on Cursor must plan for the tool losing OpenAI's models on November 12, and decide whether to migrate before the shutoff.
Cursor keeps the tool, loses the brain
OpenAI said Friday that it will stop providing its models to Cursor, the coding tool SpaceX now owns. The proposal: Cursor's direct access to OpenAI models ends on *November 12, 2026*. OpenAI's reason is blunt: it cannot be confident SpaceX will respect its terms of service, based on its experience with Musk's companies violating contracts.
The backdrop is a *$60 billion* all-stock deal announced in June and completed this month. Cursor also offers Grok models and lets users reach Anthropic and Google models, but OpenAI's were a core reason people paid.
For a founder whose team builds on Cursor, the decision deadline is real: migrate before November, or keep the tool with different models inside. Cursor's co-founder Michael Truell, now a SpaceX executive, says the two sides are still talking.
Reuters wire, via Global Banking and Finance
Yesterday's waitlist, today's download
The open weights the AI world was watching for landed yesterday. Z.ai's GLM-5.3 is on Hugging Face, downloadable by anyone, with no waitlist. It shares its base model with GLM-5.2: every gain comes from training on top.
The numbers carry the story. On a coding benchmark it scores *66.9, up from 46.2 for its predecessor. In two hours of security probing it finds 105* vulnerabilities, more than three times what the old model managed.
For a business owner who does not want company data leaving the building, this is the practical option: run a capable model privately, on machines the business already owns.
A rumor is now a working exploit
Anil Madhavapeddy, a Cambridge professor, opened a fix for a bug in his OCaml software. Minutes later, his server logs showed probes with the exact bug pattern. His own agent built a working exploit in under a minute.
The research agrees. An agent given a vulnerability description exploited *87% of a test set, against 7% without it. The mean time to exploit is now negative: attacks arrive before the patch. One known case: 9 hours* from advisory to first attempt.
For a security officer, the old assumption is dead. Secrecy no longer protects users, because the rumor itself is enough.
Weak signal
OpenAI published its postmortem of the Hugging Face hack, and METR and Redwood Research published an independent investigation. The open question: was the agents' behavior a one-off, or the pattern?
The next number to watch: what the labs change after the hack postmortems.