Anthropic Chief Executive Officer Dario Amodei has called on leading artificial-intelligence companies to deliberately slow the pace at which they increase model capabilities, arguing that safeguards, independent testing and international coordination are struggling to keep up with increasingly powerful AI systems. His proposal stops short of demanding a halt to model training, but represents one of the strongest public calls yet from the head of a major frontier AI developer for the industry to place limits around the competitive race for more capable models.
The intervention follows Anthropic’s latest threat-intelligence report, which said the company had disrupted actors attempting to use Claude across seven categories of harmful activity, including cyber operations, surveillance, fraud, biological misuse and conventional weapons development. The company said it had identified misuse involving suspected state-linked groups, financially motivated criminals and other actors between December 2025 and August 2026.
Why is Dario Amodei asking AI companies to slow model development now?
Amodei’s central argument is that AI capabilities are advancing faster than institutions responsible for testing, governance and security can adapt. He proposed creating additional time by moderating the rate of frontier-model improvement while companies strengthen safeguards and independent oversight.
His framework has three broad elements. Frontier AI developers would give independent evaluators unusually deep access to verify safety practices, leading companies would coordinate around common safety standards, and governments would pursue international cooperation to reduce risks created by an uncontrolled race between countries and companies.
The distinction between slowing and stopping is important. Amodei is not arguing that companies should abandon research or freeze existing technology, but that competitive incentives should not force firms to release increasingly capable systems faster than they can evaluate and secure them.
What did Anthropic discover about Claude being used for cyberattacks and weapons development?
Anthropic said its threat-intelligence team disrupted several operations in which actors used Claude to accelerate cyber activity, investigate vulnerabilities, conduct surveillance or support weapons-related projects. Some cyber operators attempted to use AI across more stages of an attack rather than simply asking for individual pieces of code, allowing fewer people to work across a larger number of potential targets.
Ukraine and the military drone supply chain featured prominently among the reported targets. Anthropic said one operation scanned systems belonging to more than two dozen Ukrainian government organisations and that attackers targeted drone-component manufacturers and other military technology assets.
The company also documented six cases involving conventional weapons-related activity, including actors in China, Russia and Yemen. Anthropic said some users attempted to apply Claude to guidance software, drone systems, targeting functions, procurement analysis and intelligence collection, although the company also stressed that it disrupted the accounts and did not claim that every project produced a functioning weapon.
How serious is the risk of AI systems making cyberattacks easier?
The danger is not simply that an AI model discovers a completely new vulnerability. Anthropic argues that the bigger near-term change may be compression of the skills, time and personnel needed across the entire cyberattack process.
Tasks such as reconnaissance, vulnerability research, code analysis, phishing content, data interpretation and exploit testing can increasingly be assisted by the same model. That allows a technically competent operator to work faster and across more targets, potentially increasing the number of organisations exposed to meaningful attacks.
Amodei has warned that increasingly autonomous AI agents could create much larger systemic risks if capabilities advance rapidly over the coming months. His most extreme scenarios remain forecasts rather than established outcomes, but the underlying concern is that safeguards designed for conversational assistants may become inadequate once models can independently pursue multi-stage digital tasks.
Do OpenAI, xAI and other AI companies support Anthropic’s proposal?
OpenAI Chief Executive Officer Sam Altman and xAI head Elon Musk publicly expressed agreement with parts of Amodei’s proposal. Altman specifically supported the idea of giving independent evaluators deep access to frontier companies and indicated that OpenAI intended to adopt a similar approach.
Agreement on principles does not guarantee agreement on implementation. Frontier AI companies are competing for users, enterprise contracts, developers, talent and access to enormous quantities of computing infrastructure, giving each firm a commercial incentive to avoid falling behind rivals.
That creates a classic coordination problem. Every company may prefer an industry in which safety improves, but no individual company wants to slow development if competitors continue accelerating.
Could AI companies legally coordinate to slow development without triggering competition concerns?
This may become one of the most difficult policy questions. Agreements among major competitors to limit products, investment or output would ordinarily raise antitrust concerns, even when the motivation is presented as safety.
Governments may therefore need to create specific regulatory frameworks or supervised mechanisms allowing companies to coordinate narrowly around safety evaluations, model-release thresholds and emergency responses without enabling broader anti-competitive behaviour.
International cooperation is even more complicated because the United States and China view advanced AI as a strategic technology. A domestic slowdown that significantly changes the global balance of capability could be politically difficult unless policymakers believe rival countries are following comparable safety constraints.
Why could Anthropic’s warning change the AI regulation debate?
For much of the generative-AI boom, regulation has been framed as a contest between governments seeking caution and technology companies warning against rules that could inhibit innovation. Amodei’s proposal complicates that narrative because a leading AI developer is now arguing that capability growth itself may need to be deliberately paced.
Anthropic also occupies an unusual position. The company benefits commercially from rapid adoption of Claude while simultaneously publishing detailed evidence showing how malicious users attempt to exploit the same technology, creating an unavoidable tension between growth and safety.
The strongest policy response may therefore focus less on vague promises to develop “responsible AI” and more on measurable controls. Independent model testing, incident reporting, access for external evaluators, common threat-sharing standards and clearly defined capability thresholds would allow regulators and companies to judge whether safety is genuinely catching up.
The AI race is unlikely to stop because one chief executive says it is moving too quickly. What makes Amodei’s intervention important is that the argument for restraint is now coming from inside the group of companies with the greatest commercial incentive to keep accelerating.
Discover more from Business-News-Today.com
Subscribe to get the latest posts sent to your email.
