Discover how GPT-6 Astra is transforming AI interactions with its advanced task automation and cybersecurity capabilities #
OpenAI began rolling out GPT‑6 Astra on 3 September 2026, touting 'computer use' that can navigate apps and complete multi-step tasks such as posting an eBay listing via voice. Its most advanced cybersecurity capabilities are being restricted to a small set of enterprise testers under its Daybreak program.
The release marks a significant shift in how people may interact with AI. Rather than simply generating instructions or code, Astra can work through browsers, spreadsheets, websites and desktop applications to complete tasks.
OpenAI has demonstrated the model carrying out workflows such as creating an eBay listing, building software and manipulating documents after receiving voice commands. But the same abilities that make Astra useful also explain why its rollout is being controlled.
OpenAI says the model has crossed its 'Critical' cybersecurity capability threshold, meaning it can identify previously unknown vulnerabilities and develop exploit chains with the right tools and access. OpenAI is subjecting those higher-risk capabilities to stronger safeguards rather than broadly exposing them to users.
How Astra Works Across Browsers and Spreadsheets #
The biggest practical change with Astra is its ability to interact directly with software instead of merely explaining how a person should use it. OpenAI's demonstrations show the model navigating applications in a way that resembles ordinary computer use.
It can work across browsers, spreadsheets and desktop programmes, fill in forms, update customer records, organise calendars and turn research into documents or emails. It can also manipulate spreadsheets, work with scientific data and operate specialist engineering software.
The eBay demonstration illustrates why this matters. Astra was shown creating an online listing through voice instructions, turning a task that normally involves opening a website, entering information and navigating several interface elements into a conversational workflow.
OpenAI says Astra can create and test websites, install and troubleshoot software and handle complex programming tasks. For businesses, the potential benefit is less about replacing one particular application and more about allowing an AI system to work across the software employees already use.
Why Astra's Cyber Capabilities Are Kept Under Lock #
Astra's cyber capabilities are considerably more sensitive. In its safety assessment, OpenAI said Astra became the first model to meet its Critical cybersecurity capability threshold.
The company reported that the model achieved 100 per cent on ExploitBench and, during internal testing, discovered two previously unknown vulnerabilities while constructing exploit chains. OpenAI said it was disclosing those vulnerabilities to the relevant maintainers.
The distinction is important: these results do not mean ordinary Astra users are being given unrestricted access to a system designed to find and exploit real-world security weaknesses. OpenAI says the benchmark results reflecting these capabilities were obtained with Daybreak Blue access rather than the default production configuration.
The company has also introduced additional safeguards around cybersecurity use as the model moves into deployment. The launch followed internal assessments that Astra's cyber capabilities had triggered OpenAI's enhanced safety requirements.
Safety Will Shape Openai's Future Releases #
The controlled rollout reflects a broader change in how OpenAI says it will release increasingly capable models.
Sam Altman has said publicly that future releases will be paced by progress in alignment and safety rather than capability alone.
Greg Brockman, meanwhile, described the Astra launch as the beginning of an 'AGI era', signalling how significant OpenAI believes the model's broader capabilities could become. That tension is central to Astra's launch.
The model is being presented as a more capable digital worker, able to navigate software and complete complicated tasks, while some of its most powerful cybersecurity functions remain behind additional controls. For users, the immediate change is straightforward: Astra can increasingly do things on a computer rather than simply tell people how to do them.
For OpenAI, however, the harder challenge is ensuring that greater computer access does not come with equally broad access to capabilities that could be misused. The controlled rollout reflects a deliberate strategy to balance innovation against the risks of powerful AI tools. © Copyright IBTimes 2026. All rights reserved.