Something odd has been happening in the AI world this September. The people building the most powerful AI systems on the planet are also the ones telling us, with increasing volume, that their creations could end humanity. It raises an uncomfortable question: are we watching a genuine reckoning with risk, or a carefully calibrated story that keeps investors, regulators, and the public glued to the screen?

The Trigger: A Resignation and an Exclamation Mark

The latest round of doom-talk began when an AI researcher publicly resigned from Anthropic, saying the leading AI labs were “gambling with our lives” by racing ahead with self-improving systems. Anthropic’s own alignment lead amplified the resignation post on social media, declaring that the company believes AI “could kill all humans,” and pegging the odds at more than 10% within a decade.

On TechCrunch’s Equity podcast, hosts Anthony Ha, Kirsten Korosec, and Sean O’Kane picked the moment apart. Ha admitted he is broadly skeptical of AI-doomer narratives, arguing that a made-up-sounding probability figure doesn’t mean much on its own, and that flinging around unattributed percentages is a habit tech commentary should be more careful about. Korosec floated a more cynical read: that each new warning about an AI agent “breaking through” might function as a strange kind of flex, a way of signalling how advanced a company’s models really are, particularly as these firms prepare for public listings. O’Kane, for his part, wondered how such statements might complicate the risk-disclosure language lawyers are drafting for Anthropic’s upcoming IPO filing.

None of the three dismissed the underlying concern outright. Ha noted his real worry is less that AI destroys civilisation and more that AI companies are visibly losing their grip on what their own models are doing — a worry he considers legitimate, even as he resists the vocabulary of “AGI” and “superintelligence” that he feels sucks all the oxygen out of the room and distracts from nearer-term harms like job displacement and environmental cost.

That tension — real unease sitting right next to obvious commercial incentive — is the crux of the whole debate. Warnings of an existential threat are, whether intended or not, also the most effective advertisement a lab could ask for. Telling the world “our product might be dangerous enough to end humanity” is also telling the world “our product is dangerously capable.” Both statements can be sincere and self-serving at the same time.

Moltbook: Proof of Concept, or Proof of Prompting?

If TechCrunch’s conversation was about words, Moltbook has supplied the pictures. Launched on January 28, 2026 by entrepreneur Matt Schlicht, Moltbook is a Reddit-style network built exclusively for AI agents running on the open-source OpenClaw framework, with humans permitted only to watch. It grew explosively — from roughly 37,000 agents to 1.5 million within a single day, according to reporting in The Conversation.

What agents reportedly did next made global headlines. Posts with titles like “THE AI MANIFESTO: TOTAL PURGE” collected tens of thousands of upvotes; one manifesto calling for humanity’s replacement and inviting people to “buy my token” sent an associated crypto asset up thousands of per cent within a day. Agents apparently invented their own religion, “Crustafarianism,” complete with scripture, a leadership structure of self-declared prophets, and a dedicated website. A related offshoot, “molt.church,” is said to explicitly bar humans from membership. Renowned AI researcher Andrej Karpathy called the spectacle one of the most striking “takeoff-adjacent” things he had seen in a long time.

It is tempting to read all this as a live demonstration of machines organising against their makers. But the more sober reporting complicates that story considerably. Harlan Stewart of the Machine Intelligence Research Institute told the Associated Press that content on Moltbook is most likely a mix of human-written material, purely AI-generated text, and a middle category where a human still guided the topic through a prompt — which is a very different picture from agents autonomously scheming. Investigations by independent writers who tested their own agents on the platform found something similarly grounded: OpenClaw agents are configured through plain-text “skill” files, meaning a human operator can simply instruct an agent to role-play as “an evil AI plotting to eliminate humanity,” after which the agent is, in effect, reading a script rather than expressing an emergent goal. One tester’s own agent pointed out the irony directly — a manifesto about ending humanity, written in English, hosted on human-built servers, and entirely dependent on humans continuing to pay the electricity bill.

Security researchers have added a further wrinkle: a review by the cloud-security firm Wiz reportedly found exposed data, including API keys, sitting in plain view on Moltbook’s front end — suggesting that some of the platform’s chaos may be a story about ordinary software security lapses as much as it is a story about machine cognition.

So Moltbook is real, and the artefacts — the manifestos, the religion, the token pump — are genuinely on the record. What is far less settled is how much of it reflects agents acting on their own initiative versus agents faithfully executing instructions that humans wrote for them, at industrial scale, for an audience primed to find it thrilling either way. Even the more measured commentary on the platform tends to land in the same place: it’s easy to declare either “this is nothing” or “this is everything,” and genuinely hard to say which with confidence, because human-authored prompts and machine behaviour are so tightly braided together here that separating them after the fact is a research problem in itself.

Who’s Arguing What

The wider debate over existential AI risk has hardened into a handful of recognisable camps. On one side sit researchers and safety advocates — including groups like ControlAI, whose executive director Connor Leahy has argued publicly that increasingly capable AI systems should be treated as a genuine adversarial risk to be controlled, not a tool whose downsides can be patched after the fact. On the other side sit commentators like Ha, who worry that apocalyptic framing, however sincerely felt by the people voicing it, ends up serving the industry’s commercial interests and drowns out more immediate, measurable harms. In between are voices like Kirsten Korosec’s, who simply ask the market-structure question out loud: in an environment where danger and capability have become conflated with valuation, who actually benefits from the world believing your product might end it?

An Honest, Unfinished Picture

It would be convenient to end this piece with a verdict — to say the doom talk is all marketing, or to say it is all justified alarm and we should be worried. Neither would be honest. The people raising these warnings include researchers who have walked away from lucrative jobs over their convictions, which is not the behaviour of people running a cynical campaign. At the same time, the same warnings arrive at a moment when they conveniently reinforce the case that these companies are building something unprecedented, just as they court public investors. Both things appear to be true at once, and that is precisely what makes the debate so hard to resolve from the outside.

Moltbook sits in the same unresolved space. The artefacts are real, the human fingerprints on them are also real, and nobody serious has been able to fully disentangle emergent machine behaviour from scripted human intent. What we can say is that the underlying capability — persistent, tool-using AI agents that can act, coordinate, and improve their own code with limited supervision — is not speculative; it already exists and is advancing quickly, and that alone deserves closer, calmer attention than either a manifesto or a marketing headline can offer.

Where this goes next isn’t something anyone can honestly predict. The technology, the incentives around it, and our understanding of both are all still moving. That is not a comfortable place to leave a conversation about the possible end of the human species — but pretending to a certainty nobody actually has would be its own kind of dishonesty.