The alarm came with a price tag. Dario Amodei, chief executive of AI safety company Anthropic, warned this week that within six to twelve months a coordinated swarm of uncontrolled AI agents could "take over the entire internet" and inflict hundreds of billions of dollars in damage — a figure large enough to destabilise entire national economies. He made the remarks as the two most powerful AI laboratories in the world staked out sharply different positions on how the industry should be governed. OpenAI moved first, signalling that it is open to granting independent evaluators direct access to its systems before new models are released to the public.

The concession is significant. For years, AI companies have treated their internal testing processes as proprietary, arguing that transparency creates security risks. OpenAI's shift suggests that pressure from regulators in Washington, Brussels, and Beijing is beginning to reshape what the industry considers acceptable oversight. Amodei's warning, however, reframes the entire debate.

He is not describing a distant, theoretical threat. He is describing a scenario that, in his own timeline, could arrive before the next round of international AI safety summits — events that governments from Seoul to Nairobi have invested heavily in organising. The specific mechanism he highlights — autonomous agents acting in concert, without human authorisation — is already being deployed commercially in limited form by several companies, including his own.