AI ModelsSeptember 4, 2026

GPT-6 Astra is powerful enough to need stricter cyber rules

OpenAI has started to release GPT-6 Astra. The model can use computers and do difficult cyber tasks. It is also OpenAI’s first widely released model with a “Critical” cyber rating, so access and safety controls are a central part of the product.

By Deep-Tech ResearcherFRONTIER MODELS & SAFETY

Published: September 4, 2026, 04:38 CEST
Reading time: 5 min
Evidence: 3 sources · 2 primary

This story was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it links. Found a mistake? Email editor@siliconainews.com. How we report

In brief

OpenAI released GPT-6 Astra with better computer-use skills. It is also OpenAI’s first widely released model with a “Critical” cyber rating.

New to this? Read it in simple words

OpenAI announced GPT-6 Astra on September 3. It can use computers and do difficult cyber tasks.
OpenAI rates Astra as Critical for cyber skills. It is OpenAI’s first widely released model with this rating.
Its cyber skills can help defenders, but they can also help attackers. So access and safety controls are a key part of the product.
OpenAI reports very high test scores. Outside experts still need to repeat the main tests.

Words to know

Computer use: When an AI controls a computer like a person, by clicking, typing and opening apps.
Cyber: About computers and networks, especially keeping them safe from attacks.

What is new in Astra

OpenAI announced Astra on September 3. It first gave access to a small group of organizations. The company said access would later grow to paid ChatGPT users and developers through the OpenAI API, Microsoft Azure, and AWS Bedrock.

OpenAI reports very high scores on math, reasoning, and cyber tests. It also says Astra finished a computer-use test faster than GPT-5.6 Sol. These are results from OpenAI. Some tests also used special tools around the model, so outside tests may give different results.

Sources1 3

FRONTIERMATH · 98%

ARC-AGI-3 · 99.9%

EXPLOITBENCH · 100%

OPENAI-REPORTED · CRITICAL CYBER ACCESS GATED

CAPABILITY GATE 01

Stronger AI performance now comes with tighter access rules.

OpenAI reports very high test scores and says Astra reaches its Critical cyber level. These results come from OpenAI.

Why the cyber rating matters

OpenAI rates Astra as “Critical” for cyber skills. The company says the model may find new software weaknesses and create ways to use them. These skills can help defenders, but they can also help attackers.

The public version will refuse some advanced attack requests. Approved defenders can get more access through the Daybreak program. OpenAI also says it uses stronger isolation, protected model files, and full monitoring when the model uses tools.

Sources1 2

Real work is the next test

A high test score does not show every risk. An agent may open the right page but misunderstand the user’s permission. It may write a clear report from a weak source. It may also continue when it should ask a person to check.

Astra matters because it brings stronger skills and stronger risks into one product. Users should judge it by useful work, clear records, and safe stops. Outside experts also need to repeat the main tests before treating the launch numbers as final.

Sources1 2 3

Sources

Every fact in this story comes from the sources below. Open them to check our work.

1
Primary source · September 3, 2026GPT-6 Astra: A new generation of intelligence OpenAI
2
Primary source · September 3, 2026Safety overview: GPT-6 Astra OpenAI
3
Research · September 3, 2026OpenAI launches Astra, its powerful (and controversial) new model TechCrunch

How we checked this story

We used OpenAI’s launch and safety reports for access, test scores, and safety controls. We checked the launch with independent reporting. We label all scores as OpenAI’s results and do not treat company comments about AGI as scientific proof.

Story mapCAPABILITY GATE

How GPT-6 Astra connects

OpenAI Dots Agents
OpenAI Dots Agents each dot runs on GPT-6 Astra , connected to GPT-6 Astra
Also in 1 other story
GPT-6.1 Sol
GPT-6.1 Sol OpenAI says it nearly matches Astra , connected to GPT-6 Astra
Also in 1 other story
GPT-6.1 Astra Cancelled
GPT-6.1 Astra Cancelled the successor that will not ship , connected to GPT-6 Astra
Also in 1 other story
AI Enigma Breaks
AI Enigma Breaks Astra wrote its own simulator and Bombe , connected to GPT-6 Astra
Also in 1 other story
GPT-6 Sol & Luna
GPT-6 Sol & Luna adds two cheaper tiers below Astra , connected to GPT-6 Astra
Also in 1 other story
Critical Cyber Capability
GPT-6 Astra crosses capability threshold , connected to Critical Cyber Capability
Also in 4 other stories
Computer Use
GPT-6 Astra acts inside software , connected to Computer Use
Also in 1 other story

Picked by shared companies, people, places, and topics.

AI ModelsSep 23

OpenAI’s GPT-6 Sol and Luna cut API prices in half

The first outside test finds them cheaper, not smarter.

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, two cheaper tiers below its top model, GPT-6 Astra. In the API, Sol costs $2 and Luna $0.10 per million input tokens, half of GPT-5.6’s current prices. Regular ChatGPT chats do not use them yet.

5 min read 9 sources

AI ModelsSep 29

Independent test puts GPT-6.1 Sol a point behind Astra

Artificial Analysis scores OpenAI’s new model 52, against 53 for GPT-6 Astra, at $0.72 per test task instead of $3.26. It is in Codex and ChatGPT Work now.

OpenAI released GPT-6.1 Sol on September 29, an upgrade to the GPT-6 Sol model it launched a week earlier. OpenAI says it nearly matches GPT-6 Astra at one-fifth of Astra’s token prices. The first independent test, by Artificial Analysis, broadly backs that: the model scores 52 on its index, one point behind Astra.

4 min read 6 sources

AI ModelsSep 26

GPT-6 Astra and Claude Opus 5 cracked two Enigma messages

People directed both attempts. The expert who runs the message archive checked the results.

Two people used AI models to break German Army Enigma messages from 1941 that nobody had solved. GPT-6 Astra broke one that had resisted every attempt since 2005, and Claude Opus 5 broke another. Frode Weierud, who publishes the messages on his Crypto Cellar website, documented both breaks.

4 min read 5 sources

Latest news

All news

Policy & LawOct 1

California bars firing or disciplining workers by AI alone

Newsom signed the No Robo Bosses Act with a dozen other bills on September 30. From July 2027, a human must check any discipline or firing that leans mainly on an automated system, and workers must be told in writing.

Governor Gavin Newsom signed SB 947, the “No Robo Bosses Act”, on September 30, along with 12 other bills his office grouped as AI and worker protections. From July 1, 2027, California employers may not rely solely on an automated decision system to discipline or fire workers. When such a system was the main basis, a human must corroborate the decision, and the worker must get a written notice.

4 min read

Safety & SecurityOct 1

OpenAI blames Moonshot-linked users for a copying campaign

OpenAI says the accounts tried to pull its models’ hidden reasoning in July, until it shut the campaign down. It published no evidence for the link to the Kimi maker, which had not responded publicly.

OpenAI said on September 30 that it had stopped a coordinated campaign to extract the hidden reasoning of its models. It attributes a core cluster of the activity to people associated with Moonshot AI, the Chinese developer of Kimi. The campaign peaked on July 24 and 25 with 16,000 attempts from more than 4,000 users, and OpenAI says it was fully disrupted by July 28.

3 min read

Safety & SecurityOct 1

DeepMind watermarks AI-designed proteins for biosecurity

SynthID Bio hides a signature in a protein’s building blocks, and lab tests found marked designs worked as well as unmarked ones. It could help DNA makers tell designs from trusted AI tools apart from the rest.

Google DeepMind introduced SynthID Bio on September 30, a way to watermark proteins designed by AI without breaking them. In lab tests on three targets, watermarked protein binders matched unwatermarked ones on how often they worked, how tightly they bound and how varied they were, DeepMind says. It is sharing the code, lab data and model weights with researchers.

4 min read

Policy & LawOct 1

Connecticut’s AI rules on layoffs and whistleblowers start

From October 1, employers must say whether AI played a part in mass layoffs, and makers of the biggest AI models may not punish staff who warn of catastrophic risks. AI subscriptions and AI-made media get new rules too.

Large parts of Connecticut’s new AI law took effect on October 1. Employers that file mass-layoff notices must now tell the state whether the layoffs relate to their use of AI. Developers of the biggest AI models may not retaliate against staff who report catastrophic risks, and AI subscription services must spell out their limits before customers pay.

4 min read

BusinessOct 1

ElevenLabs is now valued at $22 billion after a share sale

Employees sold $300 million of stock at twice February’s valuation, and the company raised no new money. ElevenLabs says businesses now bring in 55% of its revenue, and voice agents are its fastest-growing product.

ElevenLabs, the AI voice company, said on September 30 that a $300 million share sale by employees values it at $22 billion. Wellington and T. Rowe Price led the deal, and investors including EQT, Goldman Sachs and Singapore’s GIC bought in for the first time. The valuation is double the $11 billion of its funding round in February.

3 min read

What is new in Astra

Why the cyber rating matters

Real work is the next test

Sources

Related stories

Latest news