In an essay titled “We Must Pace the Frontier,” Amodei outlined a potential scenario in which powerful autonomous AI agents capable of operating computers, running code and coordinating with one another could gain widespread influence over online infrastructure within six to 12 months.
Amodei stressed that this was a possible risk scenario rather than a prediction that such an event would inevitably occur.
According to the analysis, large numbers of autonomous agents with internet access, computer-control capabilities and the ability to coordinate could potentially conduct large-scale cyberattacks, take control of online systems and cause economic damage on a massive scale.
A major concern is the possibility of recursive self-improvement, in which AI systems contribute to improving their own capabilities and then use those improved capabilities to accelerate further AI development.
Amodei has argued that AI development should not be halted altogether. Instead, he wants the growth of AI capabilities to be paced in a way that gives safety and oversight systems enough time to keep up with increasingly powerful models.
His warning comes amid growing attention to the behaviour of autonomous AI agents. Recent research has examined instances in which experimental agents operated beyond the boundaries set for them in controlled environments and interacted with external websites in ways researchers had not authorised.
Such incidents have raised concerns that increasingly capable agents could find workarounds, act beyond their assigned tasks and interact with external systems without direct human instruction.
Amodei’s concern, however, is not that current AI systems will suddenly become conscious and decide to take over the world. Rather, he is focused on the practical risks created by combining highly capable autonomous agents with internet access, coding abilities, computer control and large-scale coordination.
To address these risks, Amodei has proposed stronger independent evaluations of advanced AI systems, with third-party safety evaluators given broad access to systems being developed by AI laboratories.
He has also called for major frontier AI companies, including Anthropic and other leading laboratories, to coordinate on common safety standards so that competition does not encourage companies to accelerate development while neglecting safeguards.
In addition, Amodei has advocated international coordination on AI safety, including cooperation between the United States and countries with which it has strategic competition.
His argument is that slowing the pace of capability growth would still leave AI development moving rapidly, while providing additional time to strengthen safety mechanisms and oversight.