Anthropic exit, then a 1-in-10 warning⁠.

Jacob Coxon, who has worked at both OpenAI and Anthropic, has resigned from Anthropic, saying AI firms are "gambling with our lives"¹. He said the labs are carelessly racing to build "superhuman systems" they cannot control³.

Hours later, a senior Anthropic safety researcher wrote in a post that there is a more than 10 percent chance that AI "could kill all humans" by the end of the decade³. Sky News reported the same figure, framed as the next decade⁴. The BBC described it as the latest in a series of increasing warnings about the safety threat posed by artificial intelligence⁵.

The Financial Times put the resignation in a longer run of departures, saying it highlights mounting unease inside Silicon Valley companies racing to build ever more powerful systems². The warnings come from inside the firms building the systems, not from outside critics².

Two things are worth separating here. One is a probability estimate about frontier systems, offered by a researcher in a public post³. The other is the fact of a resignation from one of the two best known labs, by someone who worked at both of them¹.

Frontier risk is not your risk register

Nothing in this changes what your invoice matching agent or your customer email triage did this morning. A 10 percent estimate about civilisational risk is a claim about the largest models being trained, made by a person with a view and no way to prove it. It is not a finding, and it is not about the narrow tools most companies in Leeds or Bristol actually run. Treat the number as a temperature reading inside two American labs rather than as data.

What it does tell you is something duller and more useful: the people closest to these systems disagree, in public, about how well they are controlled. That is a supplier governance question. Ask who at your vendor can turn a model off, what changes without warning when a model is updated, and which of your processes would break if it did. No UK regulator has commented in the material we have seen today, so there is nothing to comply with yet, only something to note.

Also today

• UK safety body left out of Anthropic testing

  Anthropic reportedly withheld access to its Mythos 5.1 model from the AI Security Institute, the first time the UK body has been left out of pre-release evaluations of an Anthropic model⁷.

• Pentagon says allies cannot keep pace

  Cameron Stanley, the Pentagon's chief digital and artificial intelligence officer, said Nato and Five Eyes partners including the UK lack America's resources, experience and scale, and that Washington is working to help them avoid its own mistakes⁶.

• DeepMind pair raise €3.2m for fusion software

  Two former Google DeepMind staff have raised €3.2m to build software for fusion start-ups, a reminder that lab alumni are now selling tools to other technical industries rather than only building models⁹.

• Investor-backed missile maker plans European factories

  Covenant, backed by Peter Thiel, plans to mass-produce long-range "deep strike" missiles from factories in the US, Germany and Israel, targeting a mid-six-figure price per Anthem missile at full production¹⁰.

• Managed services firm hires three channel leads

  GCX Managed Services has appointed new channel leaders for the UK, Italy and the DACH region as the managed network and security provider pushes for growth through partners⁸.