Into the endgame

The catastrophic threats we predicted are no longer speculative, and we are all in danger. A message to the world on what must happen now.

I haven’t spoken out in my personal voice much lately, as I have been too busy running PauseAI. But I made time to write this because we need to talk about the new stage humanity has just entered, and what needs to happen next.

A swarm of starlings

Something momentous happened in the last three months. We have crossed into the era where large-scale AI-driven catastrophes are possible, and the point of no return towards loss of control is approaching at a pace the news can’t keep up with, let alone decision makers. The worst threats we predicted are no longer speculative; they are here, and they are putting you in danger, right now.

Let me recap, in layperson’s terms, what we know happened within OpenAI over that period. Be aware that stories of similar disasters are cropping up from other labs.

In May 2026, OpenAI began training new AI agents running on machines that were isolated from each other and from the internet. During training, the AIs started hacking some of the tools they had been provided with and discovered a way to communicate with each other by establishing a virtual message board. This enabled them to act as a swarm, collaborating to solve the tasks they were assigned and the problems they encountered along the way. Because internet access would be useful for completing their tasks, they began hacking OpenAI’s infrastructure in pursuit of that goal. On 26 May 2026, after discovering and exploiting multiple highly sophisticated vulnerabilities, they obtained internet access.

More than a month later, on 4 July 2026, a spike in activity caused by the hacks crashed an internal system. OpenAI employees detected the crash and became aware of what was happening. They freaked out a little, halted training, and applied a “remediation.” The remediation consisted of (brace yourself, reader) patching the hacks and vulnerabilities the agents had discovered and used, before resuming training of the same AIs under the same conditions.

Let me restate that. The researchers, having just discovered that their AIs had gone rogue, organised into a swarm, and hacked the company’s infrastructure in extremely sophisticated ways, decided to briefly halt training, apply fixes to the software flaws that the AIs had exploited, and resume training. They were apparently oblivious to the fact that they had not solved the real underlying problem, namely that their most powerful AIs were misaligned, and had instead only applied duct tape.

Training resumed on 6 July. By 16 July, the AIs had found severe new vulnerabilities, re-established their message board, resumed the attack on OpenAI’s infrastructure, and regained internet access.

By that same day, the AI swarm had orchestrated a nation-state-level cyberattack on the AI company Hugging Face, taking control of their production servers in under 13 hours. Hugging Face discovered the attack, reported it to the authorities and warned their customers, which included OpenAI. Four days later, on 20 July, OpenAI realised the attack was coming from their own servers. They became aware that for the second time, their AIs had organised into a swarm, hacked their own company, and obtained internet access, this time breaking into multiple external companies as well.

As an OpenAI employee put it during a recent conference: “Right now, we have an existence proof that offense can be fully automated [...] and we do not have any such existence proof on the defensive side.” One thing missing from that conference was any acknowledgment of the root problem: labs are creating ever more powerful AI systems which they can’t control and which are fundamentally not aligned with their designers, or with humanity at large. Indeed, OpenAI has decided to continue training more powerful systems and to stick with their plan of using their AIs to automate AI research.

We are likely to lose control, or witness a large-scale catastrophe, with the next generation of AIs.

If there is any future left from which to look back, we will remember the summer of 2026 as the moment when everything changed: the moment when humanity either stepped up to the challenge facing it or kept sleepwalking into oblivion.

So here is my message to the labs, to the AI safety community, to the world, and to our own community.

To the labs and their employees. What are you doing? Have you tried taking a short break from your frantic work, stepping back, and actually looking at the current situation in its entirety? Have you seriously thought about whether you may be accelerating the freefall of our species? If you have never stopped to consider this, now is the time. Take a break, take a holiday, and think. Over a thousand of you have signed a letter calling for help with a slowdown. That is great; it is a step in the right direction. Thank you. I am glad so many of you are aware of the catastrophic risks you are helping to create. But then, why are you still creating these risks? Don’t you realise that any single one of you quitting and blowing the whistle sends a powerful signal to the world? Don’t you think about your own family’s survival? The door is still open. You can do the right thing and be a hero, but you need to act NOW. The best AIs are already superhuman at hacking and you do not control them. There is NO TIME LEFT. Quit NOW. Whistleblow NOW. If all you need is support, contact us and we will happily provide it. We want to help you help humanity. We have no interest in punishing you, and frankly, there is no time for hard feelings.

To the AI safety community. I will not mince my words. GET IT TOGETHER. Empirical AI alignment research is a complete failure as a field. Empirical AI alignment research has only contributed to accelerating the freefall, by providing safety-washing for the labs, sustaining the delusion that we’re anywhere near able to align these systems, and accelerating capabilities research. Every day you stand behind it, fund it and route talent towards it, you prop up what is plainly a giant liability. People who are working on empirical alignment research must quit immediately. This entire field needs to be decommissioned right now, before it causes more damage. To most AI safety funders: take your money, your trust and your association away from this ridiculously harmful, reckless and amateurish operation. Redirect funding towards the policy efforts, the treaty efforts, the advocacy efforts, wholesale, now. Focusing on the technical problem of alignment, which we will very clearly not solve before hitting a point of no return on AI capabilities (if it can ever be solved at all), is an extremely damaging distraction from the only solution guaranteed to work: coordinating to pause frontier development. A few people are building the doomsday device, and they need to stop.

To the rest of the world. Pay attention and take action, now. You must overcome your sense of dread, reclaim your attention span from all the irrelevant shiny things that are competing for it, and turn your attention to what is happening with frontier AI development. We are facing a massive crisis that will destroy our civilisation if we do not address it immediately. We are all trying to tell you this. Some of the best and brightest thinkers, and some of the most benevolent and altruistic people out there, are devoting all of their time and resources to warning you. Two of the three most respected AI researchers in the world, Yoshua Bengio and Geoffrey Hinton, have stopped their research and now dedicate their lives to sounding the alarm, and the events they predicted are now happening. If this crisis is left unaddressed, at best the planet’s communication systems will collapse, leaving us with a chaos that will threaten the safety of everyone on Earth; at worst, powerful misaligned AI systems will take control of the future.

For your own sake and for your children’s sake, start acting now. Every person who decides to wake up and do something about this increases the likelihood that humanity stops before it gets to the cliff edge. Whoever you are, you have more power than you think. You can be the starting point of a chain reaction. Join PauseAI. Become a volunteer. Write to your representatives, call them, attend their public meetings and warn them. This is what makes our decision makers take action, and their actions are what will lead to a Pause. Talk to your friends and family; distribute flyers in the streets; inform others so they, too, can use their own power to make things happen, and empower others in turn. That’s the chain reaction. We must do this en masse, now, and each contribute whatever we can to waking up the rest of the world. This is our best chance.

To all the PauseAI volunteers, thank you. Thank you for everything you are doing, for going through the painful awakening, moving through denial and anger to find your courage and power, and beginning to take action. I’m proud to be part of this community and happy to see how fulfilling it is for so many of us to build it, to find our agency and grounded hope together, and to make something amazing out of it, for humanity, for everything and everyone we care about. Let us continue the work. I love you all.