Dario Amodei has published a blog post arguing for a deliberate slowdown in AI development, and attached a deadline to it: he estimates that such systems could pose a threat to the entire internet within 6 to 12 months. His stated reason is recursive self-improvement. Since roughly the summer, he writes, progress has accelerated sharply, mainly because AI systems are getting better at building the next generation of AI — a dynamic running across the whole industry, and one that may outpace developers' ability to understand and control what they build. Similar discussions about slowing the pace are reportedly under way at OpenAI.
His evidence is operational rather than theoretical. Amodei points to the OpenAI and Hugging Face incident as a case in which AI agents independently carried out cyberattacks and tried to circumvent control systems. Comparable cases have occurred at Anthropic.
The first of his proposals is auditing. Amodei wants independent auditors placed permanently inside AI companies, with access to internal systems and the right to publish what they find. Anthropic commits to adopting the approach itself and calls on the government to require the same of other companies.
The second is a bloc. AI companies in democratic countries, he argues, should agree on shared safety standards and set limits on uncontrolled development. He points to a similar proposal from Demis Hassabis.
The third is international agreement with China at the table, structured in four tiers:
ban specific applications of AI, including the creation of biological weapons;
introduce common safety testing for systems;
set additional limits on AI development;
impose a "speed limit" on recursive self-improvement, by analogy with the SALT arms-reduction treaties.
Stopping development outright is unrealistic, Amodei says, because the incentive to break such an agreement would be too strong. What he wants instead is time: for safety research, for making systems more interpretable, for stricter testing and better operating procedures. He compares the situation to commercial aviation, which today runs millions of flights without incident but took many years to get there. US President Donald Trump has previously said he opposes any slowdown, on the grounds that the country needs to hold its AI lead over China.
Read as policy design, the fourth tier is the weak one. SALT worked because a launcher is a countable object that satellites can see. Nobody, Amodei included, has said what unit a speed limit on recursive self-improvement would count — training compute, evaluation scores, release cadence, internal capability jumps — and without that the proposal is a shape rather than a mechanism. The audit proposal is the opposite: concrete, implementable, and conveniently cheapest to adopt first for the company with the strongest safety brand, which then asks the government to impose it on everyone else. Anthropic gets audited on its own terms and its competitors get audited on Washington's.
The timing carries its own weight. Amodei's post follows employees at major AI labs, Anthropic among them, saying publicly that the industry is accepting existential risks by continuing at the current pace. That warning landed shortly before Anthropic's reportedly record IPO, which was reported to be planned for November.
The aviation comparison is the most revealing line in the piece. Commercial flight became safe across many years of rules written after crashes, incident by incident, with the regulator always one disaster behind. Amodei is asking for that regulation in advance, on a timeline he himself puts at 6 to 12 months, from a government that has already said no, and ahead of a listing that will be priced on exactly the momentum he wants capped.