An Anthropic Exit Sharpens the Debate Over AI Extinction Risk
AI researcher Jacob Coxon has resigned from Anthropic, saying leading AI companies are "gambling with our lives." TechCrunch
Coxon had worked at OpenAI before joining Anthropic. TechCrunch reported his exit on September 13, 2026, as part of a wider look at renewed warnings of catastrophic AI risk.
Anthropic's alignment lead, the official in charge of keeping AI systems under human control, issued a separate warning. He said, "We really do earnestly believe AI could kill all humans!" He put his personal estimate at more than a 10% chance within the next decade.
For timing, OpenAI had released a model called Astra a few weeks before September 13, 2026. Coxon's resignation followed that release and added to questions about how labs test, limit and ship very capable models, often called frontier systems.
Inside-lab warnings drive the debate
TechCrunch's Equity podcast discussed the dispute in an episode with Kirsten Korosec, Sean O'Kane and Anthony Ha. Equity is described as TechCrunch's flagship podcast about the business of startups. The show is produced by Theresa Loconsolo and edited by Kell.
That episode was recorded before Anthropic CEO Dario Amodei published his plan for more cautious AI development. The hosts were responding to the resignation and the public risk estimates, not to Amodei's proposal.
TechCrunch lists Equity alongside Build Mode, hosted by Startup Battlefield Editor Isabelle Johannessen with episodes every Thursday, and StrictlyVC Download, hosted by Editor-in-Chief Connie Loizos with Alex Gove. TechCrunch
The two Anthropic-linked statements differ. Coxon spoke as a departing researcher criticizing industry risk-taking. The alignment lead spoke as a current safety official, giving a number and a ten-year period for extinction risk. One is a staff exit that points to internal disagreement. The other is a forecast about loss of control.
Parallel warnings from Geneva and Basel
On September 7, 2026, UN human rights chief Volker Turk warned that artificial intelligence could pose an existential risk to humanity. Reuters Turk pledged to press AI firms to reduce AI risks.
The head of the Bank for International Settlements said in September 2026 that the AI boom poses new financial stability risks. Reuters The bank estimated the world's five largest technology firms will invest more than $1 trillion in AI between 2025 and 2026.
Those warnings cover different risks. Turk addressed harm to people and rights. The BIS addressed borrowing, concentration of power and very large capital spending on computing at scale. Both came from outside the labs and pointed to gaps in oversight, not to any single model release.
The broader context here is that insider dissent now travels faster than formal oversight. A resignation letter, a probability shared on social media, and a podcast debate can set the agenda before a CEO paper or a regulator statement lands. Engineers feel that shift directly, because choices about release, safety tests and incident response stay inside companies while public expectations are set outside them.
Looking at what this means for technical teams, the questions are narrower than the headlines. What ideas support a 10% estimate for ten years. What tests would prove it wrong. What release controls, access rules and monitoring after launch would lower it. What public detail would let outside researchers check the work. None of these questions require accepting the estimate. They require making the thinking open to checking.
In my view, the BIS figure belongs in the same engineering discussion. More than $1 trillion across five firms in two years concentrates computers, talent and infrastructure choices. That focus shapes which safety work gets money, which public tests survive, and which failures get attention. Money risk and control risk are not the same problem, but both grow when scale increases without matching independent checks.
It is worth flagging the limits of what is public. The confirmed record holds resignations, stated probabilities, a model name and release window, a podcast debate, a coming caution plan from Amodei, and two institutional warnings. It does not hold test results, system cards, or details of Amodei's plan. Readers should treat the probability as the author's belief, not as a measured rate.
The longer view here is that this could lead to better safety engineering. Clearer tests before launch, better limits for systems that can improve themselves, shared reports on failures, and buyer demand for checks that can be proved would help products whether extinction estimates end up high or low. My own children grew up while seat belts, spam filters and app permissions quietly became normal. The protections that lasted were built into the system, not argued in the abstract. AI safety will likely follow that path, from public warning to quiet and reliable controls.
