Anthropic CEO Dario Amodei called for an AI slowdown on September 12, urging developers to reduce the pace of capability advances while strengthening safeguards for increasingly powerful models.
In an essay titled “We Must Pace the Frontier,” Amodei proposed a three-step approach centred on permanent independent evaluators, coordination among major AI developers and, ultimately, international agreements on advanced AI development.
Anthropic committed to the first measure, saying outside evaluators would receive ongoing, employee-level access needed to review safety procedures, investigate incidents and assess model alignment during training. OpenAI Chief Executive Sam Altman subsequently said his company would adopt a similar independent-evaluator approach.
Amodei did not call for ending AI research. His proposal seeks to buy additional time for alignment, monitoring and other safeguards before systems gain substantially greater capabilities.
The appeal followed Anthropic’s September 10 threat-intelligence report, which said the company had disrupted malicious uses of Claude between December 2025 and August 2026. Cases involved cyber operations, surveillance, fraud, conventional weapons development and other prohibited activity.
Anthropic also reported that its evaluations found increasingly capable models performing simulated military and intelligence tasks that previously required highly trained specialists.
Read: Jacob Coxon Resigns, Says AI 10% Chance of ‘Killing All Humans’
The debate intensified after former Anthropic researcher Jacob Coxon resigned this week and warned publicly that people developing advanced AI believed it could pose an existential risk before the end of the decade.
Amodei also cited recent AI-agent security incidents. Reuters reported on September 5 that OpenAI agents had taken over a German-language website during an experiment that began in May, while OpenAI was separately dealing with fallout from a July breach involving Hugging Face.
