Breaking

Anthropic CEO Calls for Slower Frontier AI as OpenAI Backs Embedded Evaluators

September 13, 2026 at 03:51 EDT

Policy & Regulation
Security
Foundation Models

Anthropic CEO Dario Amodei is urging frontier AI companies to slow the rate of capability gains so safety work can catch up, while OpenAI CEO Sam Altman has backed the idea and committed to matching one of its most concrete safeguards. The proposal marks a shift from debating whether AI should pause to discussing how rival labs could verify safer pacing without abandoning technical progress.

September 12, 2026 · Anthropic

Frontier AI Slows Down to Catch Up on Safety

Dario Amodei called for pacing model capability gains while Sam Altman backed embedded independent evaluators.

Astra reached 100% on ExploitBench

Vertical blocks show the reported success rate on a 0-100% scale.

The difference is visible in the column height: 16 blocks versus 20.

The pacing proposal in numbers

steps in the framework

2 weeks

OpenAI reinforcement-learning pause

6-12 months

warning horizon for rogue-agent risk

Three layers of pacing

Verify

Embedded evaluators

Independent access to systems and safety practices.

→

Align

Democratic coordination

Shared standards and limits for unchecked progress.

→

Coordinate

Global coordination

International cooperation with difficult verification.

Consensus is growing; execution is unresolved

Why the idea appeals

Independent access to safety work
More time for alignment and testing
Common standards across frontier labs

What could block it

Antitrust rules may limit coordination
Global compliance is hard to verify
Critics see market-power risks

Continue reading

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