Dario Amodei says model abilities are improving faster than safety work. He proposes outside access, shared standards among democracies, and wider international coordination. His timelines are warnings, not proven forecasts.
New to this? Read it in simple words
- On September 12, Anthropic CEO Dario Amodei published a plan to slow AI progress. He says AI abilities are improving faster than safety work.
- His plan has three parts. They are outside reviews, shared rules among democracies and wider talks with other governments.
- He says shared rules are needed, because one company or country may not slow down. Anthropic itself competes in the market these rules would cover.
- He warns agents might take over large parts of the internet in six to twelve months. This is a worry, not a measured prediction.
- AI agent
- An AI that takes steps on its own to finish a task.
The plan tries to slow capability without stopping all work
Amodei published his proposal on September 12. He argues that better models may soon help create even better models. He believes this could make progress too fast for researchers, companies, and governments to manage safely.
His first step is deeper outside review. Anthropic says it will give selected evaluators access similar to employees. Those experts could inspect safety work and report serious incidents while protecting private and dangerous information.
The second step is shared rules among democratic countries. The third is wider talks with other governments. Amodei says common controls are needed because one company or country may keep racing when others slow down.
Dario Amodei proposes outside access and national and international standards. Important triggers and enforcement details remain open.
The strongest claim is also the least certain
Amodei warns that a group of future agents might take over large parts of the internet within six to twelve months. He presents this as a worry based on recent progress and known security incidents. It is not a measured prediction with a public probability.
Associated Press and Axios confirmed the new call for slower development. They also report support from leaders at other AI companies. Support in public does not yet show which model release, training run, or business plan those companies would delay.
There is also a conflict of interest. Anthropic competes in the same market that it wants to regulate. Its technical experience is useful, but public rules should not give one company power to slow smaller rivals or hide weak evidence.
A real pacing system needs visible triggers
A useful rule should name the capability that triggers action. It could use tested ability in cyberattacks, biological work, self-copying, or control of many tools. The test method and important limits should be open to qualified review.
The response should also be clear. A failed test might lead to more safeguards, limited access, a delayed release, or a pause in larger training. Independent reviewers should check whether the company completed the required work before moving again.
The proposal moves the debate from general fear toward possible controls. Its value will depend on details that are still missing. Readers should watch for written standards, outside access, public incident reports, and proof that Anthropic accepts the rules itself.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2Research · September 12, 2026Anthropic CEO says the AI industry needs to give safety measures time to catch up Associated Press
- 3
We used Amodei’s essay for the plan and his risk claims. We used Associated Press and Axios to confirm the timing and outside reaction. We label the short timeline as his warning, not a fact or a forecast proven by data.