> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Anthropic's Safety Team Puts Human Extinction at 10 Percent
- URL: https://wire.fourthweb.ai/anthropics-safety-team-puts-human-extinction-at-10-percent/
- Published: 2026-09-09T09:56:28.000Z
- Updated: 2026-09-09T11:30:43.000Z
- Description: The lab that promised "Constitutional AI" just watched two of its own researchers publicly declare the odds of human extinction are worse than Russian roulette.
- Author: Travis Wright
- Tags: AI Agent Economy, AI Agents, AI Infrastructure, AI Governance, OpenAI, Anthropic, Funding Rounds

**The lab that promised "Constitutional AI" just watched two of its own researchers publicly declare the odds of human extinction are worse than Russian roulette.**

### The Summary

- [Jacob Coxon resigned from Anthropic](https://www.theverge.com/ai-artificial-intelligence/991927/anthropic-ai-kill-all-humans?ref=wire.fourthweb.ai), accusing both [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) and [OpenAI](https://wire.fourthweb.ai/tag/openai/) of "racing straight to self-improving superintelligence and gambling with our lives"
- Hours later, [a senior Anthropic safety researcher put the odds of AI killing all humans by 2030 at over 10 percent](https://www.theverge.com/ai-artificial-intelligence/991927/anthropic-ai-kill-all-humans?ref=wire.fourthweb.ai)
- Coxon previously trained systems at OpenAI before joining Anthropic, suggesting the safety dysfunction spans the entire frontier AI industry
- The warning comes from inside the company literally named after its commitment to AI safety

### The Signal

The math here is brutal. A one-in-ten chance of extinction is orders of magnitude worse than any risk humanity currently accepts. We don't build bridges with a 10 percent collapse rate. We don't approve drugs with a 10 percent fatality rate. Yet [Anthropic researchers now openly estimate those odds for AI-driven human extinction within five years](https://www.theverge.com/ai-artificial-intelligence/991927/anthropic-ai-kill-all-humans?ref=wire.fourthweb.ai).

What makes this different from the usual AI doomerism is the source. This isn't a philosophy professor or a longtermist blog. These are the people actually training the models. [Coxon worked at both OpenAI and Anthropic](https://www.theverge.com/ai-artificial-intelligence/991927/anthropic-ai-kill-all-humans?ref=wire.fourthweb.ai), meaning he's seen the safety culture at the two labs racing hardest toward AGI. His resignation isn't a prediction. It's a performance review.

> "Racing straight to self-improving superintelligence and gambling with our lives."

The timing matters. Anthropic just raised billions at a $60 billion valuation based partly on its safety-first positioning. The company was founded by former OpenAI researchers who left specifically because they thought OpenAI was moving too fast. Now their own people are saying Anthropic has become the thing it was created to oppose.

The phrase "self-improving superintelligence" is doing heavy lifting here. That's not GPT-5 writing better emails. That's a system that can recursively upgrade its own architecture faster than humans can track, let alone control. [The researchers worry about "superhuman systems" they cannot control](https://www.theverge.com/ai-artificial-intelligence/991927/anthropic-ai-kill-all-humans?ref=wire.fourthweb.ai), which suggests the current safety frameworks assume human-level or near-human capability. Once models cross into truly superhuman territory, the assumption breaks.

Here's the ugly economics: the AI labs are trapped. They've raised too much money at too high a valuation to slow down. Investors expect exponential capability growth. The market rewards the fastest mover. Safety work is expensive, slows deployment, and generates no revenue. Every dollar spent on alignment is a dollar not spent on training runs that might unlock the next capability jump.

### The Implication

If you're building on top of frontier models, you're gambling that the people training those models have better safety measures than their own researchers think they do. That's a hell of a supply chain risk. The agent economy only works if the underlying intelligence doesn't slip the leash.

Watch for two things: whether other Anthropic researchers follow Coxon out the door, and whether any of the labs actually slow down. If neither happens, you have your answer about how seriously the industry takes its own warnings. The 10 percent odds aren't a call to action. They're just the number everyone has quietly agreed to accept.

### Sources

[The Verge AI](https://www.theverge.com/ai-artificial-intelligence/991927/anthropic-ai-kill-all-humans?ref=wire.fourthweb.ai)