In a September 8–9 window, researcher Jacob Coxon publicly announced he was leaving Anthropic and accused frontier labs of racing toward self-improving superintelligence while “gambling with our lives.” Anthropic Alignment Science lead Evan Hubinger replied that the extinction worry is earnestly held, that his personal estimate is greater than 10% within a decade, and that Anthropic does not yet have a clear plan for superintelligence alignment — while also saying risk from current models is low.

Anthropic researcher raises alarm after quitting job: AI could kill us all · CBS News

Quick Take

Jacob Coxon publicly said he resigned from Anthropic and accused frontier labs of racing toward self-improving superintelligence while gambling with human lives. Evan Hubinger, Anthropic’s Alignment Science lead, replied that the extinction concern is earnestly held inside the company, that his personal estimate is greater than 10% within a decade, and that Anthropic lacks a clear superintelligence-alignment plan — while describing risk from current models as low.

  • Confirmed: Resignation announcement + Hubinger reply as reported by major outlets Sep 8–9.
  • Opinion: Extinction timelines and “gambling” language belong to named people.
  • Not proven: A new Anthropic board policy, a product recall, or an official corporate probability.

What the video shows

Embed: “Anthropic researcher raises alarm after quitting job: AI could kill us all” (YouTube nNtAKglFxbk), CBS News, September 9, 2026. House note: network packaging of the resignation and Hubinger reply. Prefer over hotter CNN titles. Still secondary to reading attributed quotes in Ars/BI/TC.

What’s new

For a general AI brief, the newsworthy cut is the alignment-lead corroboration of a departing researcher’s warning — not a new Claude release. That pairing is what moved the story from one person’s exit thread into mainstream coverage on September 9.

Evidence

Press secondaries (Ars, BI, TechCrunch, CBS). Consistent on Coxon’s resignation posts and Hubinger’s reply, including the >10% personal estimate and the “no plan yet for SI alignment” caveat. BI notes labs did not immediately comment in some coverage.

Still missing. An Anthropic or OpenAI formal statement adopting a numeric extinction probability; independent verification of private “fear” claims beyond the public posts.

What this does not prove

  • It does not prove Hubinger’s >10% line is Anthropic company policy. Attribute as personal opinion.
  • It does not prove current models are about to cause human extinction. Hubinger described current-model risk as low; the worry is future SI / RSI.
  • It does not prove a product recall or board policy change. This is an insider speech-act story.

Why it matters

If you ship agents or buy frontier APIs, treat this as a governance signal, not a product changelog. Named insiders say the existential-risk conversation is earnest and that alignment for superintelligence is unsolved. That is important. It is still not the same as evidence that today’s tools are about to cause human extinction. Keep the fence: attribute, date, and separate opinion from policy.

What to watch next

  1. Any formal Anthropic or OpenAI statement adopting (or rejecting) a numeric extinction probability.
  2. Follow-up from named alignment leads clarifying personal vs policy language.
  3. Whether product or access policies change — absence of comment is not agreement.

Bottom Line

Coxon’s resignation warning and Hubinger’s personal >10% extinction estimate are real attributed speech acts — not Anthropic company policy and not proof today’s models are about to cause human extinction. Attribute, date, and fence current-vs-future risk.

Sources