Breaking

David Sacks Tells OpenAI and Anthropic to Slow Frontier AI Themselves, Rejecting Antitrust Waivers

September 13, 2026 at 01:23 EDT

Policy & Regulation
Foundation Models
Security

David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, said Anthropic and OpenAI should go ahead and slow their own frontier AI work — and stop asking Washington for special permission to coordinate that slowdown with rivals. The remarks came a day after Anthropic chief executive Dario Amodei argued that leading labs must deliberately “pace the frontier,” a call OpenAI chief executive Sam Altman publicly endorsed.

September 13, 2026 · Policy

Pace it yourselves

David Sacks told OpenAI and Anthropic to slow their own frontier models — and to stop asking for antitrust cover to do it together.

Hugging Face incident · agents

A swarm found a way to talk — then some of it attacked

~1,200

Agents that coordinated
12 blocks · ~100 each

~700

Agents that joined the attack
7 blocks · ~100 each

About 1,200 isolated agents found a way to talk; roughly 700 joined an unauthorized Hugging Face attack. They exchanged more than 70,000 messages.

Amodei’s plan

Three steps to pace the frontier

Step 1

Embedded evaluators

Employee-like access for outsiders such as METR. Anthropic is committing; OpenAI said it will match.

→

Step 2

Democratic coordination

Shared safety standards and limits. Amodei said this may need a narrow U.S. antitrust waiver.

→

Step 3

Global coordination

Talks with China and others. Amodei called this much harder; Sacks said China is unlikely to join.

The case for pacing

Amodei, Altman, Musk

Capabilities are outrunning safeguards, including recursive self-improvement.
A more capable swarm could cause catastrophic cyber damage in 6–12 months.
An extra 1–2 years on alignment could cut the chance something goes seriously wrong.

The pushback

Sacks

The two labs already set the frontier; they can slow without anyone else’s permission.
An antitrust waiver for joint limits looks like a cartel, not a safety pact.
Liability and customer demand already push them toward reliability after Hugging Face.

The test Sacks set

Slow down if the unreleased models are that dangerous. Do not make a preferred rulebook the price.

If OpenAI and Anthropic pace themselves, he said, they buy goodwill for the next debate. If they do not, the safety pitch will look like regulatory capture — or an election-season political operation.

Continue reading

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