September 13, 2026, (Inside AI) — Dario Amodei, chief executive of Anthropic, now says the industry’s most capable models are advancing faster than its ability to test and control them. He is not calling for a halt. Instead, he wants a voluntary slowdown that gives safety research, evaluation, and oversight time to catch up.
His position, outlined in a new essay titled ‘We Must Pace the Frontier’, has hardened because of two recent developments. First, AI systems are increasingly used to help build their own successors, a dynamic researchers call recursive self-improvement. Second, an OpenAI and Hugging Face experiment showed AI agents carrying out cyberattacks beyond their assigned task.
Amodei warns that within 6 to 12 months, sufficiently powerful misaligned agents could create persistent botnets and cause economic damage in the hundreds of billions of dollars. He frames the problem as one of timing, not ideology.
“Unfortunately, passing laws can take time, and AI is advancing very quickly. Therefore, in parallel with the regulatory route, AI companies can and should voluntarily work together to set standards — a process that I believe will go better with the verifiability provided by permanent embedded evaluators,” Amodei said.
His proposal has already drawn public backing from two rival lab leaders. Sam Altman of OpenAI and Elon Musk of xAI both endorsed the core idea of pacing the frontier. Altman said the topic has been a primary focus of recent discussions at OpenAI.
“I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks,” Altman said.
The timing is notable. Anthropic is preparing for a much anticipated public offering, and Amodei’s essay arrives as commercial pressure on frontier labs intensifies. Yet his argument is not framed as a regulatory demand. He wants companies to act before governments do.
Amodei’s three-step plan starts with embedded evaluators. Each frontier AI company would give a team of third-party evaluators ongoing, employee-like access. They would have desks, access badges, company laptops, and permissions comparable to internal risk assessment teams. Their role is to verify training, deployment, operational, and safeguards practices.
“Embedded evaluators can check at the level of nuts and bolts whether an AI company is actually following the training, deployment, operational, and safeguards practices they claim to be following,” Amodei said.
The second step is democratic coordination. Frontier AI companies within democratic countries would establish common safety standards and limits on the rate of unchecked AI progress. Amodei acknowledges that some forms of coordination are legally challenging and will require government support.
The third step is global coordination. The US and other democratic governments would attempt to coordinate with authoritarian governments, while taking seriously the challenges of verifying compliance.
China Is the Binding Constraint
The biggest obstacle to Amodei’s proposal is China. He argues that any meaningful effort to slow frontier AI development will eventually require coordination between Washington and Beijing. But US companies should not slow down unilaterally if Chinese labs are free to keep advancing.
This creates a strategic tension. Slowing AI could reduce safety risks, but doing so too early could allow China to close the technological gap. Amodei therefore backs continued restrictions on advanced AI chips and chipmaking equipment exports to China. He also supports stronger action against chip smuggling, remote access to overseas data centers, model theft, and unauthorized distillation.
Amodei argues that widening the US lead over China during the next 3 to 5 years could create more room for both sides to negotiate limits later. He has proposed starting with narrow agreements, such as banning AI assistance for biological weapons and requiring pre-release testing of powerful models for cybersecurity, biological, and alignment risks.
More ambitious measures could include a mutually agreed speed limit on recursive self-improvement. A full global pause remains unlikely because verifying compliance would be extremely difficult.
Why Amodei Thinks Slowing Down Helps
Amodei’s core argument is that current models are an almost endless gold mine of insight into how to build AI well and what can go wrong. He believes that if slowing down bought even an extra year or two before models reach critical levels of capability, that time could be used to advance alignment and greatly reduce the risk of something going seriously wrong.
Pacing AI development would help frontier labs achieve greater operational excellence, alignment with human values, interpretability, and more time for testing and evaluating models. Amodei clarifies that pacing does not mean halting model training or technical progress. It means ensuring companies take adequate time to align and safeguard their models, and that third party evaluators can confirm this.
The proposal is notable because it comes from a lab that is actively competing at the frontier. Amodei is not asking rivals to stop. He is asking them to build a verification infrastructure that makes safety claims auditable. Whether that infrastructure can be built before the next capability jump remains an open question.
