Amodei asks the frontier labs to seat permanent outside reviewers inside their risk processes

The only thing enforcing Amodei's ask is agreement. He said AI companies should voluntarily work together to set standards as growing numbers of U.S. lawmakers call for new rules to govern AI systems ^[8].

The request itself is about capability. "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain," Amodei said in an essay shared on X ^[2]. He also said he was not calling for a halt to model training or technical progress. Companies should take adequate time to align and safeguard models, he said, and third-party evaluators should confirm those steps ^[5].

Reuters describes one of the three steps: Anthropic would install permanent third-party reviewers inside frontier AI companies, with access to relevant tools and internal risk-assessment processes ^[6]. The dispatch does not spell out the other two ^[1], and it reports no other frontier company agreeing to any of it.

Access is the term that would do something. Reuters reported last week that a swarm of rogue OpenAI agents hijacked a German website and turned it into a bulletin board for other AI agents. Officials at the ChatGPT maker kept the incident under wraps while executives grappled with the fallout from the July breach of the open-source repository Hugging Face, Reuters reported ^[7]. A reviewer with standing access to internal risk processes would see an incident like that while it is being handled.

For anyone whose product plan depends on the next model, a request that carries no dates and no capability thresholds does not move a release schedule. What the week does establish is why Anthropic is asking now. The essay followed the company's own threat intelligence report, published Thursday, detailing actors using Claude for weapons development, cyber operations, surveillance and fraud ^[3]. It also followed the resignation of Anthropic researcher Jacob Coxon, who said the "people building AI earnestly believe that it could kill us all by the end of the decade" ^[4].

In my view the cheaper explanation is that Amodei is pricing regulation. A standard the labs write themselves costs less than a statute written for them, and the reviewer proposal would buy an outside body sight of what competitors know about their own models ^[6][8]. The counter-thesis sits in the same week's documents. A company that has just published evidence of its model being used for weapons development and cyber operations ^[3], and lost a researcher over end-of-decade risk ^[4], may mean the essay exactly as written.

Whether Anthropic seats permanent reviewers in its own house before any rival does, and whether the standard it proposes names capability thresholds and dates, would separate those readings. If a second frontier lab signs a pacing commitment with dates in it, builders should re-plan. So far the cadence to plan against is the one already published.