Anthropic researcher resigns and warns AI companies are outrunning safety

Jacob Coxon announced on X on Tuesday, 8 September, that he had resigned from US artificial intelligence developer Anthropic, citing fears about the industry's pursuit of systems that improve themselves, according to [Axios](https://www.axios.com/2026/09/09/openai-artificial-general-intelligence-safety). His announcement brings a researcher’s objection to the pace of development into public view.

Updated: 9 September 2026, 14:22 CEST.

Coxon, a 27-year-old British researcher, worked on model pretraining at OpenAI before joining Anthropic this year, [AFP reported](https://www.tvanouvelles.ca/2026/09/09/elles-jouent-avec-nos-vies-apres-openai-un-chercheur-en-ia-quitte-anthropic). He accused both companies of proceeding irresponsibly and said he was leaving the industry. AFP reported that neither company immediately responded to its requests for comment.

In his [statement, reproduced by Thread Navigator](https://threadnavigator.com/thread/2097476196791709843/), Coxon argued that commercial rivalry was driving decisions with consequences extending far beyond individual companies. He described Anthropic as understanding the stakes but feeling compelled to finish ahead of competitors. He urged researchers to demand different conditions for further development and advocated coordination between laboratories, with government intervention if necessary.

Anthropic safety researcher Evan Hubinger also expressed concern, according to AFP. He put his personal estimate of the risk of human extinction from AI above 10% over the coming decade, while describing the risk from current models as low. That figure is his assessment; the reporting does not establish it as a measured probability or scientific consensus.

Anthropic’s published position provides a different perspective on continuing development. In its [February safety-policy revision](https://www.anthropic.com/news/responsible-scaling-policy-v3), the company described a voluntary framework intended to reduce catastrophic risks and strengthen accountability. Its [policy register](https://www.anthropic.com/responsible-scaling-policy) records subsequent revisions and an August risk report. These documents establish what Anthropic says its safeguards involve; they do not independently demonstrate their effectiveness.

For Belgium, the direct connection is European oversight. The [European Commission](https://digital-strategy.ec.europa.eu/fr/policies/guidelines-gpai-providers) says its enforcement powers over general-purpose AI providers took effect on 2 August 2026. Its rules place additional obligations on providers of models presenting systemic risks.

The next identifiable checkpoints include Anthropic’s own safety commitments: its [published roadmap](https://www.anthropic.com/responsible-scaling-policy/roadmap) sets a 30 September target for a phase of security research. Whether those commitments answer Coxon’s objections remains unresolved. The sources reviewed provide no direct corporate response to his resignation beyond AFP’s account of unanswered requests.