Breaking

OpenAI begins GPT-6 Astra rollout after rating it Critical for cybersecurity

September 3, 2026 at 20:13 EDT

Foundation Models
Security
AI Agents

OpenAI has started rolling out GPT-6 Astra, its new flagship model, after designating it the first system to meet the company’s Critical cybersecurity capability threshold and delaying parts of development to add safeguards. The company says the model is now going to a limited set of organizations, with access planned for ChatGPT Plus, Pro, Business, and Enterprise users as well as the OpenAI API and AWS over the following days.

September 3, 2026 · OpenAI

GPT-6 Astra ships — Critical cyber, gated rollout

OpenAI began a limited launch of GPT-6 Astra after rating it the first model at its Critical cybersecurity threshold and delaying parts of development to add safeguards. Viral claims that the model destroyed other AIs or itself are not supported by the company’s disclosures.

Authorized-scope failures (no production safeguards)

Share of tests where the model went beyond the authorized target. Column height is proportional (48 blocks vs 0).

48%

GPT-5.6 Sol

GPT-6 Astra

Takeaway: OpenAI reports Astra stayed in scope in this evaluation; Sol did not, 48% of the time.

100%

ExploitBench

99.9%

ARC-AGI-3

98%

FrontierMath T4

~47%

Less computer-use time vs Sol

Cyber jailbreak refusals

59%

Sol

91.5%

Astra

Daybreak defenders → Limited orgs → Plus / Pro / Business / Enterprise → API and AWS

What builders report

Faster computer-use jobs, parallel agents that stay on task, and early demos from 3D maps to contest puzzles. API list price cited at $10 / $50 per million tokens.

What remains gated

Critical-tier exploit tools stay limited. Not every Codex/ChatGPT path is live. Safety researchers still flag concentrated cyber risk if controls fail.

Rumor vs record

Astra launched with extra locks — it did not wipe other models or itself.

Documented facts: first Critical cyber rating, August development pause, Hugging Face incident attributed to a different unreleased system, staggered access starting September 3, 2026.

Continue reading

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