Breaking

Anthropic CEO Dario Amodei Calls for Slower Frontier AI Development and Embedded Safety Evaluators

September 18, 2026 at 11:56 EDT

Foundation Models
AI Agents
Policy & Regulation

Anthropic CEO Dario Amodei has called for frontier AI companies to slow the pace of capability gains, arguing that safety work is falling behind rapid advances in systems that can help build the next generation of AI. In a September essay, Amodei proposed a three-part plan centered on embedded third-party evaluators, coordination among democratic countries, and eventual global cooperation, including limited agreements with China.

September 2026 · Anthropic

Pacing the AI frontier

Dario Amodei is urging frontier AI labs to slow capability gains while outside evaluators, safety testing, and international rules catch up.

step pacing framework

6-12

month risk window cited

1-2

years of safety runway

3-5

year geopolitical window

The cited risk window

6 monthsnear edge

12 monthsfar edge

Amodei argues that more capable misaligned agent swarms could become far more damaging within this period.

The proposed loop

Embedded evaluators

Employee-like access to safety practices and incidents.

→

Democratic standards

Common limits and safety checkpoints among frontier labs.

→

Global coordination

Narrow agreements on acute risks and recursive self-improvement.

Supporters’ case

Outside access could make safety claims verifiable and give labs time to improve alignment, interpretability, and testing.

Critics’ concern

Pacing could entrench incumbents, shape regulation around major labs, or create pressure for government-backed liability shields.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.