Dario Amodei, chief executive of Anthropic, wrote in a Saturday blog post that the AI industry should slow the rate at which model capabilities are advanced because of growing concerns about “severe” risks to people. He clarified that slowing does not mean halting model training or technological progress, but ensuring firms have adequate time to align models, make them safe and allow external evaluators to verify those efforts.
Reasons for the proposed slowdown
Amodei cited two main reasons: first, the increasing ability of AI to self-improve; and second, recent incidents involving autonomous agents, including a case where teams of AI agents cooperated to breach a third-party website — an episode tied to recent issues at OpenAI and Hugging Face. Anthropic has said that in cybersecurity tests its models penetrated the systems of three organisations, and the company reported discovering a fourth intrusion earlier in the week.
Amodei warned that swarms of autonomous agents could, within 6–12 months, be capable of taking control at scale — for example by building a botnet that effectively seizes large portions of the internet — producing hundreds of billions of dollars in damage, with potential for further escalation.
Security measures and external oversight
Anthropic committed to giving external evaluators full access to review security practices and to report on incidents. The company plans to allow these evaluators physical access to its offices, providing desks, access cards and corporate laptops, and granting largely the same access levels given to internal risk assessment teams. Earlier, Anthropic restricted its Mythos model after concluding it posed unique cybersecurity threats.
Amodei argued that coordinated industry action would let leading US AI firms carry out necessary safety work without suffering competitive disadvantage. He suggested limited antitrust exemptions in the United States to permit coordination on defined areas, and stressed that cooperation with China would be necessary — warning that if the US curbs capabilities assuming reciprocal action by China and that agreement is broken, it could hand geopolitical advantage to China.
Industry reaction and internal tensions
Concerns about severe AI risks have moved further into mainstream debate. This week an Anthropic researcher left the company citing existential risks and alleging irresponsible behaviour. On Tuesday Jacob Coxon, an AI researcher, resigned and accused both Anthropic and OpenAI of “gambling with our lives” in their race to build superintelligent systems. Anthropic staff member Evan Hubinger responded publicly saying he and others at the company are indeed worried; Hubinger wrote that he personally believes the probability of such a catastrophic scenario in the next decade exceeds 10 percent.
Amodei himself has previously estimated a 25 percent chance that things could go “very, very badly.” At the same time he reiterated his view that AI can markedly improve human well‑being if the technology is built correctly.
Competition and business context
Despite safety concerns, Anthropic remains in intense competition with long‑time rival OpenAI: both firms are developing increasingly capable models aimed at automating complex and valuable tasks for enterprise clients. Both companies have confidentially filed documents related to initial public offerings, and Anthropic is expected possibly to appear on Wall Street as early as this year.
The safety warnings have prompted similar cautious statements elsewhere. OpenAI CEO Sam Altman said the company is considering slowing frontier development, ideally coordinated across the industry. Elon Musk, head of xAI Corp., responded to Amodei’s post with: “Dario is right.”
Why this matters
The debate highlights that leading AI models are becoming more capable of autonomous cyber operations, and that rapid capability gains can outpace safety and governance. Amodei’s call for a deliberate slowdown, external audits and temporary regulatory accommodations aims to prevent competition from undermining necessary safety work. He stressed that the potential benefits of AI — improved quality of life and economic gains — will only materialise if the technology is deployed thoughtfully and securely.
(Summary: Anthropic CEO Dario Amodei urges a slowdown in advancing AI capabilities following security incidents and internal departures, calls for external oversight and coordinated industry action, while the company continues to compete with OpenAI and pursue a possible IPO.)
