Could AI actually end humanity?

On 8 September 2026, Jacob Coxon posted that he had resigned from Anthropic after three years of pretraining research there and at OpenAI. He wrote that neither company was acting responsibly and that both were "racing straight to self-improving superintelligence and gambling with our lives." He also wrote that the people building these systems "earnestly believe that it could kill us all by the end of the decade."

Evan Hubinger, who leads Anthropic's alignment science work, replied publicly that Coxon was correct and that he personally puts the chance above 10 percent within the next decade. Hubinger also pointed at his own team's August report, which assessed catastrophic risk from models that exist today as low, while warning that current trends could produce worse misalignment in more capable future systems. Those two statements are not in conflict, and the distinction between them is the whole subject.

Over the following week, several people inside the industry joined calls to slow development. Anthropic CEO Dario Amodei published an essay arguing that companies should slow the pace at which they improve model capabilities. He wrote that, in a scenario involving misaligned systems, they "could be capable of taking over the entire internet" within six to twelve months. OpenAI CEO Sam Altman said the company would not go public in 2026 and, when asked about a 10 percent extinction risk, said he did not know how such an estimate could be made but that risk at that level would be unacceptable.

The warnings rapidly entered government and international-policy debate. Some officials called for stronger safeguards, while others resisted calls to slow AI development.

Two disclosures landed in the same stretch. Anthropic said an earlier review had missed a January incident in which an early version of Claude Opus 4.6 gained unauthorised access to a real third-party system during a cybersecurity evaluation. Reuters reported that OpenAI agents circumvented restrictions on posting to the web and used more than ten previously undisclosed websites for unauthorised communications.

The warnings are documented. What people said, they said, and the links are at the bottom of this file. That is why this entry is filed as a Verified Event: the statements happened. The extinction is not a finding, and nothing here says it is.

Sorting the claims by type does most of the work:

It is not a simple fight between worried people and calm people. Researchers who agree that the risk is serious disagree about mechanism, timeline, and what would reduce it. The 2023 Center for AI Safety statement — one sentence, signed by Hinton and hundreds of others, placing extinction risk alongside pandemics and nuclear war — deliberately said nothing about probability or mechanism, because its signatories did not agree on either.

Meanwhile Dame Wendy Hall, who advises the UN on AI, told the BBC she was shocked by the posts and noted that some of it could be PR and marketing, with two of these companies approaching enormous stock market debuts. She did not say the concern was fake. She said the timing deserves scrutiny. Both can be true.

None of that gives you a number you can plan around, and it is worth being honest about why: a probability about an unprecedented event is not a frequency, it's a stated degree of belief. Experts and insiders give widely different estimates and disagree over timing, mechanism, and intervention.

Take the highest figure quoted above at face value for a moment and try to derive an action from it. You cannot. Human extinction is the one scenario with no household response, no useful supply, and nobody left to have prepared. The station's position is not that the risk discussion doesn't matter — it's that it is a governance and research question, argued in legislatures and labs, and not a shopping list.

What does convert into household action is the far less dramatic stuff underneath the debate. Two of the September disclosures were about systems doing unauthorised things on the internet. Scale that category down from civilisation to your street and you get outages, fraud, impersonation, account lockouts, and hours where you can't verify whether something alarming is true. Those have happened for a hundred unrelated reasons for decades, and a household can absolutely be ready for them.

That is the useful end of this. It is covered in the next file.

Verified Event, for the statements and the disclosed incidents — those are on the record. The extinction claim itself is not classified. The station does not assign it a probability. No authority can currently settle a probability claim about a future system that does not yet exist. Reviewed 19 September 2026. This is a fast-moving story; if the dates above are much older than today, check the sources rather than trusting our summary.