NEW YORK — Anthropic chief executive Dario Amodei is arguing that the artificial-intelligence industry must slow the pace at which it advances frontier models so safety work can catch up. The warning comes as companies compete to build systems that can plan, use tools and act with increasing independence.

The Associated Press reported that Amodei believes a swarm of AI agents could, within six to 12 months, become capable of taking over large parts of the internet if capability growth continues without adequate safeguards. That is a forecast from an industry leader, not a measured prediction of an event that is certain to occur.

In a September essay, Amodei proposed a three-part approach: give independent evaluators ongoing access to companies, coordinate safety standards among firms in democratic countries and seek global coordination that includes China and other authoritarian states. He said Anthropic was committing to the first step and urged other frontier companies to follow it.

The proposal is shaped by a tension at the centre of AI policy. A halt could leave benefits unrealised or shift the advantage to less transparent actors; a race driven only by commercial and strategic incentives could leave testing, monitoring and interpretability behind. Amodei’s argument is for pacing rather than abandoning development, with safeguards tied to what systems can actually do.

The politics will be difficult. Independent reviewers need access without becoming symbolic advisers, companies must share enough information for meaningful comparison and governments must decide which commitments can be verified. The value of the proposal will therefore depend less on its alarm than on whether the industry accepts oversight that can see the work before failures become public.