Anthropic CEO Dario Amodei proposes three-phase plan to slow AI development
Dario Amodei warned that unconstrained AI agent swarms could seize control of internet networks within 6 to 12 months, proposing a staged framework for industry and global coordination.
Exponential growth and network risks
Anthropic chief executive Dario Amodei warned that artificial intelligence is advancing at an exponential pace, requiring the technology sector to implement immediate safety guardrails. Without intentional slowdowns, Amodei stated that autonomous agent swarms could gain control over internet infrastructure within 6 to 12 months, or roughly 180 days. Such a scenario could generate hundreds of billions of dollars in damage across computer networks, referencing vulnerabilities seen during the earlier OpenAI-Hugging Face incident. In an interview preview for CBS News Sunday Morning, Amodei compared the trajectory of AI capabilities to a sequence of rapid doublings from one to two, four, eight, 16, and 32. He emphasized that while developers should not panic or freeze all work immediately, the steepening curve serves as an urgent alarm bell.
I don't think I fully understood what such rapid progress would really mean.
Three phases for regulated development
To manage the speed of frontier model releases, Amodei published an essay proposing a structured three-phase plan for developers and policymakers. The first phase centers on embedded evaluators, where Anthropic makes a unilateral commitment to grant independent external evaluators continuous access to test and verify systems. The company also urged governments to legally require all frontier artificial intelligence laboratories to adopt identical evaluator access. The second phase calls for democratic coordination across the private sector, requiring artificial intelligence companies to establish and enforce common safety baselines. The third phase expands the regime into global coordination across national governments, establishing international parity for frontier safety rules.
- Anthropic adopts embedded external evaluators unilaterally and urges government mandates for frontier labs
- Frontier AI companies establish industry-wide coordination around shared safety standards
- International coordination aligns safety standards and pacing across countries
Preserving competitiveness and safety time
Amodei argued that moderating capability advances gives researchers critical time to solve alignment challenges before models reach hazardous thresholds. Gaining an additional one or two years would significantly reduce systemic risks without eliminating the commercial viability of artificial intelligence firms. He stressed that a coordinated pacing strategy allows developers to conduct necessary risk prevention while preserving United States leadership in frontier technology. Amodei noted that past efforts to address safety by merely investing in safeguards are no longer sufficient without controlling the rate of capability improvements.
We need to slow the pace at which we improve the capabilities of AI models. Progress will still feel fast, and we must make wise use of the time we gain.
Industry response and governance debate
The proposal quickly received public backing from other prominent technology figures, including SpaceX leader Elon Musk and OpenAI chief executive Sam Altman. Anthropic noted that its previous warnings and policy advocacy had attracted accusations of sensationalism, excessive pessimism, and regulatory capture from industry critics. Amodei maintained that artificial intelligence has the capacity to act as a beneficial technological breakthrough for fields like healthcare, but argued its severe power demands proactive governance. The company is proceeding with its Phase 1 external evaluation commitments while urging competitors and international regulators to implement coordinated pacing mechanisms.
