Anthropic defends safety record as AI risk warnings surface

Anthropic has pushed back against criticism from a former researcher who accused the firm of "gambling with our lives". The comments come as the company, known for its Claude models, prepares for a potential public listing.

The researcher, Jacob Coxon, resigned and claimed both Anthropic and its former employer OpenAI were racing toward self‑improving superintelligence without adequate safeguards. He warned that such a race could endanger humanity.

In response, an Anthropic spokesperson said the company has "some of the strongest safeguards in the industry" and has long been open about AI’s dual‑use nature.

We have always been transparent that AI will bring both enormous benefits and unprecedented risks.

The firm highlighted its testing regime for dangerous capabilities, covering areas such as cybersecurity and biology, and cited a responsible scaling policy that dictates when development should pause or safeguards should tighten.

Internal concerns and external pressure

Evan Hubinger, who leads alignment science at the company, warned that the risk of AI causing human extinction could exceed ten per cent within a decade if superintelligence emerges through recursive self‑improvement. He stressed that current models pose a low risk, but future systems could be different.

Samuel Marks, Anthropic’s scalable oversight lead, said developers themselves sometimes believe the technology could lead to human extinction, and that commercial pressure fuels continued advancement.

Regulatory backdrop and political reaction

Across Westminster, the debate has intensified. Darren Jones called for a multinational treaty to govern superintelligence development, writing to Andy Burnham, António Guterres and Mathias Cormann that governments must step in.

The UK’s AI Security Institute complained it was not given early access to the latest Claude model, Claude Mythos 5.1. A Cabinet Office spokesperson said the institute continues to work closely with developers, noting that AI risks cross borders.

Meanwhile, Matt Clifford announced he will step down as chair of the Advanced Research and Invention Agency after criticism over potential conflicts with his new role at Anthropic.

What comes next

Analysts expect the safety narrative to become a central theme in any IPO prospectus, with investors demanding clear risk‑mitigation strategies. Regulators may also move toward formal frameworks, especially if industry players push for coordinated safeguards.

For a deeper look at the researcher’s departure, see the report on the resigning scientist.