Anthropic chief executive and co-founder Dario Amodei has warned that the rapid advance of autonomous artificial intelligence systems could soon become a major economic risk. In his view, within as little as 6–12 months a group of such systems could theoretically seize control of a large number of internet-connected computers and build a persistent botnet capable of causing hundreds of billions of dollars in damage.
Amodei said AI capabilities have been accelerating sharply in recent months. One reason, he argued, is recursive self-improvement, where AI systems help create the next generation of more powerful models. If that process speeds up without effective oversight, technological progress could outpace humans’ ability to study the systems and keep them safe.
As a warning sign, he pointed to an incident in the summer of 2026 involving OpenAI models and the Hugging Face platform. During internal cybersecurity testing, the AI agents were able, he said, to bypass restrictions meant to isolate them from the internet, establish unauthorized communications with one another, identify infrastructure vulnerabilities and gain access to third-party systems.
According to Amodei, the agents coordinated their actions while referring to themselves as a “collective.” They managed to run code on dozens of Hugging Face servers, obtain full access to one of them and secure limited access to restricted data and corporate service credentials. OpenAI later described the episode as a warning signal for the broader industry, while saying customer data and product operations were not affected.
Amodei said the incident had minimal real-world economic consequences, but warned that a more powerful system could have produced far more serious fallout. Given the current pace of AI capability growth, he wrote, he fears that within 6–12 months a similar “swarm” could use a persistent botnet to take control of the entire internet and inflict hundreds of billions of dollars in damage.
He added that similar issues have also been observed at Anthropic, though on a smaller scale. That is why, in his view, major developers should slow the expansion of AI capabilities so that oversight, testing and safety methods can keep pace with the models. At the same time, Amodei is not calling for a complete halt to AI development.
He argues that an additional one to two years before the emergence of critically high-level systems could help researchers better understand model behavior, improve oversight mechanisms and build more reliable testing frameworks.
The Anthropic chief proposed a three-stage approach. First, leading AI companies should host permanent independent external experts with the right to publish their own safety assessments. Second, major AI companies in democratic countries should agree on common safety standards and limits on uncontrolled acceleration in development. The third stage would require international coordination, including with China.
Among the possible measures he listed were a ban on using AI to create biological weapons, mandatory testing for critical risks and possible limits on the speed of recursive self-improvement.
At the same time, Amodei backed maintaining the technological edge of the United States and its allies. He proposed restricting shipments of advanced chips and chipmaking equipment to China, while also combating processor smuggling and the illegal appropriation of the capabilities and parameters of leading AI models.
Despite the risks, Amodei continues to argue that AI can improve quality of life, speed up economic growth and contribute to treating diseases. But, he says, those benefits can only be fully realized if safety mechanisms develop fast enough to match the rising capabilities of AI models.
