Breaking

OpenAI's Altman calls U.S. Astra review productive as government scrutiny of frontier AI grows

September 3, 2026 at 06:27 EDT

Policy & Regulation
Foundation Models
Security

OpenAI chief executive Sam Altman said the Trump administration's voluntary review of the company's forthcoming Astra model was a "productive process," arguing that closer work with governments will matter more as frontier systems grow more capable. Speaking at the G20 Innovation Ministerial in Chapel Hill, North Carolina, Altman confirmed that officials had examined Astra even though participation is voluntary: "But we of course did it." He called Astra "a significant step forward."

Sept. 3, 2026 · OpenAI · G20 Chapel Hill

Astra hits Critical. Washington already looked.

Sam Altman called the Trump administration's voluntary review of OpenAI's forthcoming Astra model a productive process, as the first Critical-rated system heads toward release with tighter government engagement.

Jailbreak refusal, to scale

Astra refused far more disallowed cyber requests in OpenAI's tests than GPT-5.6 Sol. Column height matches the rate.

59%

GPT-5.6 Sol

91.5%

Astra

Takeaway: refusals rose from 59% to 91.5% before OpenAI cleared a Critical-rated model for release.

Cyber rating

First

Critical threshold at OpenAI

ExploitBench

100%

Known-vuln exploit score

Zero-days found

Chained in internal tests

Gov. preview

30 days

Voluntary closed-model window

Path to a limited launch

1 · Rate

Astra designated Critical for cyber capability

→

2 · Review

Voluntary U.S. preview of a closed model

→

3 · Limit

Advanced cyber tools held for testers

→

4 · Release

Broader Astra launch, still with monitors

Why some back it

Altman called the U.S. Astra review productive and said closer work with safety institutes in the U.S., U.K., and elsewhere will matter more as capability rises.

Why others worry

The framework is unpublished, open models are excluded, and critics say opaque rules can favor giant labs. Extra monitors may also pause legitimate user work.

Still missing

Congress has not set the basic framework Altman expected after his 2023 testimony.

No public thresholds Review still voluntary More capable models coming

Continue reading

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