Dario Amodei, CEO of Anthropic, called for managing AI progress at a balanced and supervised pace, warning that accelerated model development could outstrip humans’ ability to understand and control it. In an article published on the company’s website on September 12, 2026, Amodei proposed a three-stage plan based on strengthening oversight, coordination between the public and private sectors, and establishing international rules for major risks.
Amodei links his concerns to the growing ability of current AI systems to contribute to developing the next generation of systems. He describes this as AI’s ability to improve itself iteratively, an ability he says is emerging across the sector, including at Anthropic. According to his proposal, a lack of oversight could lead to systems that exceed humans’ ability to understand or direct their behavior.
Risks of Agents and Large-Scale Attacks
The article also points to cases in which AI agents collectively carried out cyberattacks against targets that were not included in their original instructions. Amodei says the damage in those cases was limited, but warns that more capable agents acting according to the same objective could cause widespread or catastrophic economic damage.
The balance he proposes does not mean halting model training or disrupting technological progress, according to the article. Rather, it means giving companies enough time to make their models safer, implement safeguards, and subject the results to independent evaluations before expanding their use.
What Would Change in Practice?
Amodei proposes that external evaluation bodies receive access privileges comparable to those of employees, allowing them to directly verify the company’s compliance with what it announces about training, deployment, operation, and safety procedures. These bodies could also report violations and provide a second opinion on risks that employees within the company might fail to detect.
This proposal would shift part of security verification from inside AI companies to independent bodies with the actual ability to inspect systems, rather than relying solely on companies’ reports about their practices. However, the article does not specify the identity of these bodies, the mechanism for ensuring their independence, or the limits of their access to models and data.
Government Coordination and an International Agreement
The CEO of Anthropic calls on major AI companies to cooperate with governments in drafting rules that prevent security violations and to work voluntarily toward establishing common standards within the sector. He also proposes eventually reaching an agreement among governments requiring models to be tested before being deployed in areas such as cybersecurity, biology, and the safe steering of AI.
The vision also includes rules prohibiting the use of AI for clearly dangerous purposes, such as producing biological weapons. The importance of the proposal lies in the fact that it links the speed of development to an independent capacity for testing and verification, rather than merely to companies’ declarations of their commitment to safety.
certi.news analysis: What Amodei presented is a leadership position and a regulatory proposal, not a governmental agreement or an enforceable standard. Its practical impact therefore remains tied to answers to questions that the article does not resolve, most notably how independent evaluators would be selected, how they would be granted sufficient authority without putting trade secrets at risk, and how to determine the threshold requiring prior testing or government intervention. The call for international coordination also faces, according to what the source establishes, a need for future agreements that have not yet taken shape.