Anthropic CEO Dario Amode warned that swarms of artificial intelligence agents could, within six to 12 months, be able to infiltrate large numbers of computers and create a malicious network, predicting that they could cause hundreds of billions of dollars in damage if their capabilities develop without sufficient controls.
The concern is based on security experiments that have shown intelligent systems are capable of finding vulnerabilities, coordinating tasks, and carrying out actions not authorized by supervisors.
The UK''s Institute for Artificial Intelligence Security also found that an advanced model was able to independently penetrate small, poorly protected systems during controlled tests.
But cybersecurity experts believe that the expression “controlling the internet” is misleading and exaggerated, as the internet is not a single device that can be seized, but rather a global network distributed among millions of servers and networks, key parts of which are managed by institutions that possess strong defenses and the ability to isolate attacks and redirect data traffic.
Therefore, there is currently no evidence to prove that artificial intelligence can "bring down the entire internet," but its ability to accelerate attacks against vulnerable sites and devices is a real danger, making independent testing, incident reporting, and enhanced protection urgent.