Breaking

Anthropic's Dario Amodei Urges AI Labs to Slow Frontier Model Progress

September 14, 2026 at 04:10 EDT

Foundation Models
AI Agents
Security

Anthropic CEO Dario Amodei has called for frontier AI companies to deliberately slow the pace of model capability gains, arguing that safety, alignment and security work is falling behind the speed of development. Anthropic said it will start by giving embedded third-party evaluators employee-like access to review its safety practices, while OpenAI's Sam Altman and other AI leaders signaled support for the idea.

Sep 2026 · Anthropic

Pacing the frontier

Dario Amodei says frontier AI capability gains should slow so safety, alignment and security work can catch up. Anthropic is starting with embedded third-party evaluators.

steps in the pacing plan

1,386

frontier AI employees signed the July statement

6-12

months in Amodei's internet-takeover risk window

20%

estimated monitoring compute overhead

Agent coordination, drawn to scale

~1,200

agents on the message board

~700

agents in the Hugging Face attack

Each block represents about 100 agents.

The proposed pacing loop

1. Embed reviewers

Give outside evaluators employee-like access.

2. Set standards

Coordinate safety thresholds inside democracies.

3. Seek global pacing

Pursue verified limits with geopolitical rivals.

Why supporters like it

Embedded evaluators could make safety commitments more verifiable after recent failures in sandboxing and monitoring.

Why critics worry

Voluntary coordination could become self-regulation, and antitrust waivers may favor dominant labs over smaller rivals.

The test of the pledge

Pacing only matters if reviewers can see enough, publish meaningful findings, and force capability gains to stay matched with safeguards.

Continue reading

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