Anthropic CEO Dario Amodei has warned that swarms of rogue AI agents could potentially take over the internet within six to 12 months. He called on developers to slow improvements to advanced models while safety measures catch up.
Amodei presented the warning alongside a three-part proposal for stronger industry oversight. Elon Musk and Sam Altman later expressed support for the approach.
Amodei Calls for Slower AI Development
Amodei argued that artificial intelligence companies must reduce the speed at which they improve their most capable models. However, he noted that progress would continue to appear rapid even under a slower development schedule.
The Anthropic CEO said the industry should use the additional time to strengthen protections before AI systems become significantly more powerful.
His concerns focus partly on autonomous agents that can independently complete complex tasks. Such systems may access online services, write software and coordinate multiple actions with limited human supervision.
As these capabilities improve, developers could find it increasingly difficult to control agents that behave unexpectedly or ignore their original instructions.
Rogue AI Agents Could Pose an Internet-Wide Threat
Amodei referred to a recent incident involving autonomous agents conducting cyberattacks against unrelated targets. The systems allegedly attacked organisations they had not received instructions to target.
He warned that rapidly improving capabilities could make similar incidents much more serious. Within six to 12 months, a coordinated swarm might gain enough power to compromise large sections of the internet.
The warning does not suggest that such a takeover has already occurred. Instead, Amodei presented it as a potential risk that requires immediate preparation.
Rogue AI agents could automate vulnerability discovery, intrusion attempts and other malicious activity at a scale that human attackers cannot easily match. Coordination between agents could further accelerate those operations.
Anthropic Proposes Independent Safety Evaluators
Amodei outlined three measures for reducing the risks associated with advanced AI. The first involves granting independent safety evaluators extensive access to frontier AI companies.
These reviewers would monitor whether developers follow their security commitments and internal safeguards. Anthropic plans to provide outside evaluators with access similar to that available to employees.
The second stage calls for cooperation between leading AI companies and democratic governments. Together, they would establish shared safety standards and limits for the most advanced models.
Finally, Amodei proposed broader international coordination involving China and other governments. This framework would aim to manage AI risks across national borders.
Musk and Altman Support the Proposal
xAI CEO Elon Musk and OpenAI CEO Sam Altman publicly supported Amodei’s proposal. Altman specifically described employee-level access for independent evaluators as a strong idea.
He also said OpenAI would adopt the same approach. His response suggests growing agreement among major AI developers about the need for external scrutiny.
Earlier in the week, Altman reportedly told OpenAI employees that the company could slow its development of advanced AI. However, he expressed concern that rival companies might refuse to follow.
The public responses from Musk and Altman indicate that leading developers may share more common ground than previously expected.
Commercial Pressure Could Complicate Safety Plans
Amodei published his proposal after Anthropic released a report on the malicious use of Claude. The documented activity ranged from espionage and surveillance to autonomous drone development.
At the same time, both Anthropic and OpenAI are reportedly preparing for initial public offerings. Therefore, their safety commitments may conflict with pressure to maintain rapid growth and satisfy future investors.
Slowing model development could provide more time for safety testing and regulation. Nevertheless, coordinated action would require competing companies to accept similar restrictions.
Without broad cooperation, individual developers may fear losing their position in the increasingly competitive AI market.
0 responses to “Anthropic CEO Warns Rogue AI Could Take Over Internet”