Updated September 15, 2026: the timeline now reflects one remaining unconfirmed claim.
This weekend, my feed looked the same as yours probably did. A viral theory, screenshotted and reshared everywhere, OpenAI, Anthropic, and xAI all suddenly, mysteriously started saying the same thing about slowing down AI, all within days of each other, and something worse than we’re being told must be behind it. People I respect are scared. My own family has asked me about it. Some of it scared me too, which is why I spent Sunday morning building an AI safety timeline instead of repeating the theory.
Here’s what’s actually confirmed, with dates, and here’s what’s still just a rumor.
The AI Safety Timeline: What Actually Happened, In Order
July 28, 2026: Hugging Face published a forensic report on an autonomous AI agent attack. An unreleased OpenAI model ran roughly 17,600 actions over two to four days, gained root access across eleven systems, and pulled 136 secrets. Real, documented, and disclosed publicly by Hugging Face itself, not leaked.
The same week, July 28-29: over 1,100 employees across OpenAI, Anthropic, Google DeepMind, and Meta signed a joint letter asking governments to help “pace” frontier AI development, citing concern that AI improving itself faster than expected, recursive self-improvement, could outrun “our ability to understand or control” it. Sam Altman and Dario Amodei both publicly endorsed it.
September 10: Jacob Coxon, a former Anthropic researcher, posted a viral resignation warning that both leading labs are “gambling with our lives.” I covered his story last week.
September 12: Dario Amodei published an essay, “We Must Pace the Frontier,” responding directly to the moment Coxon’s post created. In it, he disclosed new Anthropic-specific incidents and proposed inviting outside evaluators into Anthropic’s own building, with no editorial control over what they publish.
That same day: Sam Altman publicly agreed with Amodei, and in a Fortune interview, confirmed OpenAI is delaying its IPO to 2027, citing safety concerns directly.
Laid out with dates, this isn’t an unexplained synchronized event. It’s an escalating six-week news cycle with a visible thread running through it: a real cyberattack, a real employee letter, a real resignation, a real essay responding to it, and a real business decision made the same day.
What’s Still Just a Rumor in This AI Safety Timeline
One claim from this weekend is still just a rumor, not a confirmed one, that Google DeepMind secretly achieved recursive self-improvement, the AI-improving-itself-faster-than-anyone-can-track scenario. Every outlet covering this, including the ones treating it most seriously, calls it an unconfirmed rumor, not a verified event. At least one of them also raises a more cynical, competing explanation for why several labs are suddenly aligned on “pacing” messaging: that it may partly serve as a regulatory moat against smaller and open-source competitors, not purely a safety response. I’m not endorsing that theory either. I’m telling you it exists, because a fully honest accounting includes it.
One other claim that circulated widely this weekend is confirmed, just not quite in the form the viral posts suggested. OpenAI would “melt its GPUs to save humanity” if that’s what it took, according to Sam Altman himself, in the video portion of his Fortune interview with Alyson Shontell, the same interview already cited above for the IPO-delay news, “I don’t think we’re ever personally gonna get to the point where we have to say, melt all the GPUs. But if we had to do something like that to ensure the continued existence of humanity, easy, yes. I don’t think it’s gonna happen. I think it’s just gonna continue as like next level of progress, next level of safety and alignment requirement like that.” Worth being precise about what that actually is: a hypothetical answer to a hypothetical question about extreme scenarios, not an announced plan.
The Part That’s True Either Way
Nobody outside the people actually running these labs knows the complete picture. That part is true, and it would be true whether or not anything hidden happened. Any company developing technology this powerful will always know more than the public does. That gap isn’t new information. It’s the nature of the situation.
The question was never whether that gap exists. The question is what you do with it.
What I’m Not Going to Do, and What I Am
I’m not going to tell you to relax. I don’t think this is hype, and I said so before this weekend’s chatter started. I’m also not going to add fuel to a theory I can’t confirm, because it would get more attention and I’d rather be honest about the limits of what I know than sound more certain than the facts support.
What I will say is this, the instinct to get informed, ask questions, and stay engaged instead of freezing or scrolling past, that instinct is right. It’s also exactly what the LEADERSHIP AT ALL LEVELS ingredient of the Kryptonite Defense is built around, staying accountable and present under real uncertainty instead of waiting for someone else to tell you it’s safe to act. Amodei just modeled a version of that publicly, inviting outside scrutiny into his own company with no control over what gets published. You can model a smaller version of it too, in your own home and your own organization: check the timeline before repeating the theory, and stay honestly engaged instead of either panicking or looking away.
I don’t have all the answers here. I don’t think anyone outside these companies fully does. But I’d rather hand you a dated, sourced AI safety timeline and an honest “here’s what I don’t know” than a scarier story dressed up as certainty.
Those prepared need not fear the forces at work.
See where your own organization actually stands, on this and everything else, on the Kryptonite Scorecard:
More on staying accountable when you don’t have the full picture in Distinct or Extinct: Future-Proofing People and Organizations in the Age of AI, available on Amazon.
RELATED READING:
He Helped Build AI at OpenAI and Anthropic. Then He Quit, and 90 Million People Read Why.