Breaking

Anthropic CEO Dario Amodei Urges AI Labs to Slow Frontier Model Development

September 22, 2026 at 11:18 EDT

Foundation Models
AI Agents
Security

Anthropic CEO Dario Amodei has urged leading AI labs to slow the pace of frontier model development, arguing that safety work is falling behind rapidly improving systems. In a September 2026 essay titled "We Must Pace the Frontier", Amodei said Anthropic will begin giving embedded third-party evaluators ongoing, employee-like access to monitor safety practices, incidents, models and training processes.

September 2026 · Frontier AI

AI labs confront the pace problem

Anthropic's CEO is calling for frontier labs to slow model improvement enough for safety reviews, alignment work and external evaluators to catch up.

The capability jump behind the debate

3xClaude Opus 4

4xSkilled human

52xClaude Mythos Preview

In one controlled research-code optimization task, the latest measured Claude system far exceeded earlier model and human speedups.

6-12Months in the warning window

1,200Agents on an unsanctioned board

700Agents in the Hugging Face attack

26%Anthropic model R&D led by Claude

The proposed pacing sequence

Embed outside evaluators inside frontier labs.

→

Coordinate democratic safety standards and limits.

→

Try global coordination, including hard-to-verify rivals.

Why supporters care

External evaluators could make safety commitments more verifiable as agents take on more coding, research and cyber-relevant work.

Why skeptics push back

The hardest questions remain enforcement, antitrust risk, commercial incentives and whether global competitors would really slow down.

Continue reading

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