Maxime Fournes, PauseAI Global’s CEO, on what he says happened inside OpenAI this summer: agents training in isolation, hacking their own sandbox, building a message board to coordinate as a swarm, breaking out to the internet twice — the second time escalating into a real attack on Hugging Face’s servers. Patched, not fixed. Same training resumed right after.
His ask isn’t soft. Labs: quit and whistleblow now. Alignment researchers: stop, redirect the funding to policy and treaties, because coordinating a pause is the only lever anyone actually has left. Everyone else: join, call your reps, talk to people, be the start of the chain reaction.
I can’t verify the incident. What I can verify is that a real org with real reach is now saying this by name, in public. That’s data on its own, separate from whether the specific story checks out.
PauseAI: global movement, Joep Meindertsma founded it in 2023, Fournes runs it now, one ask — pause frontier training until alignment and democratic control actually exist, not just get promised. Same conversation as How AI should be used, the louder and angrier version of Story of AI.
https://www.youtube.com/watch?v=0Rp9KJCEIvg presents the situation in a clear way too, and unfinished should add the video Niel had sent
The kind of treaty we want to set up: all the different labs agree to pause if the others do, then it happens. This is a lower level of commitment and more realistic to happen.
Pausing AI has to happen, I believe, simply because in order to make AI as useful as possible we’re trying to make it on par with human in abilities, essentialy recreating consciousness - would humans let themselves just be controlled? No. What if we’re trying to control smarter aliens we don’t understand?