Breaking

Anthropic's Amodei Pushes Banking-Style Oversight for Frontier AI Labs

September 23, 2026 at 00:21 EDT

Policy & Regulation
Security
AI Agents

Anthropic CEO Dario Amodei has escalated the AI safety debate with a proposal for embedded third-party evaluators inside frontier labs, arguing that the fastest AI developers need oversight closer to banking supervision than conventional product review. The proposal comes as Anthropic says Claude is already leading 26% of its AI research and development work and participating in more than 90%, figures that sharpen the case for outside monitoring of systems increasingly used to build their successors.

September 2026 · Anthropic

Banking-style oversight for frontier AI

Dario Amodei wants embedded outside evaluators inside AI labs as Claude takes a larger role in building future models and agent incidents raise the cost of weak controls.

Claude's role in AI R&D

AI leads measured R&D

AI collaborates or leads

The taller column shows how much of Anthropic's measured R&D now has Claude at least collaborating under human direction.

30,000

research and engineering agents

0.002%

August agent decisions blocked

6% / 12%

AI R&D compute safety snapshots

The proposed oversight loop

Embed reviewers

Give outside evaluators employee-like access.

→

Verify practices

Inspect models, pipelines and incidents.

→

Report findings

Publish key results with narrow redactions.

Support case

Continuous review could make pacing verifiable and give the public more visibility into frontier-lab safety practices.

Open questions

No major lab has named a specific delayed model, and coordinated restraint faces antitrust and credibility tests.

Continue reading

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