Amodei says AI swarm could seize the internet in months

Dario Amodei wrote on September 12 that an AI swarm could run a persistent, internet-wide botnet within 6 to 12 months. That damage could reach hundreds of billions of dollars, and it's Anthropic's own CEO naming the number, not a hypothetical outside critic.

Alongside the warning, Anthropic made a unilateral commitment, the first step of a three-part slowdown plan. Third-party AI safety evaluators get permanent, employee-level access to the company's systems, and they keep publication rights. Anthropic can redact only for security or legal reasons, not to shape the findings.

Anthropic's own threat reporting already documents criminal groups running agent swarms in the wild. That backdrop makes the warning harder to dismiss as positioning, per VentureBeat's report.

It's also a bet against itself. Handing outside evaluators standing access and publication rights means accepting that some findings will be embarrassing, on a timeline Anthropic doesn't control.

The essay doesn't say what happens if evaluators find the risk sooner than 12 months out. It also doesn't say whether the plan's other two steps carry the same weight as this first one.