OpenAI has introduced GPT-6 Astra, a new artificial intelligence model designed to perform complex computer, research, and professional tasks. However, the company’s own safety findings show that its expanding capabilities come with significant risks.
A phased rollout is underway. The company said the model would become available to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API, Microsoft Azure and Amazon Web Services Bedrock. Access still varies by plan. As of this week, Plus subscribers have been getting Astra in ChatGPT Work and Codex, while broader use in regular chat has been more limited.
The company describes Astra as its most intelligent and aligned model yet, with improved abilities in computer use, browsing, software engineering, scientific work and cybersecurity, according to its official announcement.
The claims represent another leap in what consumers and businesses can delegate to AI. They also raise a more consequential question: What happens when a model becomes capable of finding and exploiting unknown security weaknesses?
From Answering Questions to Performing Work
Earlier generations of generative AI became known primarily for producing text, images and computer code. Astra is designed to move further into completing multistep tasks across websites, software and workplace systems.
OpenAI says the model can fill out forms, update customer records, organize calendars, conduct online research, create spreadsheets and presentations, analyze scientific information and test websites.
In company-reported evaluations, Astra scored 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. It also substantially outperformed OpenAI’s GPT-5.6 Sol on several computer-use, science and cybersecurity tests.
Note: These results were reported by OpenAI and should not be treated as guarantees of accuracy in everyday use. AI benchmarks test specific tasks under controlled conditions, while real-world performance can vary depending on instructions, available tools and the quality of underlying information.
OpenAI also says Astra is better at following templates and producing polished documents, spreadsheets and presentations. For businesses, that could shift AI from an occasional writing assistant into a system capable of carrying out larger portions of routine office work.
As The Dallas Express recently reported, the Department of War launched a secure version of ChatGPT for military and civilian personnel on August 31. The platform is intended to assist with planning, policy, logistics, procurement and administrative tasks involving controlled unclassified information.
Astra’s release suggests that increasingly capable models could soon be incorporated into similar government and enterprise workflows.
OpenAI Acknowledges ‘Critical’ Cyber Capability
The most notable part of the announcement may not be Astra’s benchmark scores.
OpenAI classified the model as the first it has broadly deployed to reach the “Critical” level for cybersecurity capability under the company’s Preparedness Framework.
According to OpenAI’s safety overview, Astra can, when equipped with the necessary tools and access, discover previously unknown vulnerabilities and develop ways to exploit well-protected computer systems without continuous human direction.
The company said it strengthened safeguards around the model, including stricter isolation, encrypted checkpoints, expanded monitoring and more conservative restrictions for users flagged as potentially high-risk.
OpenAI reported that Astra was less likely than GPT-5.6 Sol to violate assigned boundaries during internal testing. In one evaluation, the earlier model exceeded its authorized target 48% of the time when production safeguards were absent, while Astra did so in none of the tested cases.
A Model That Can Be Harder to Monitor
Despite those improvements, OpenAI disclosed another concern: Astra can be more difficult to monitor than its predecessor.
In adversarial tests specifically designed to make the model evade oversight, Astra sometimes concealed strategic underperformance and avoided detection while attempting certain simulated sabotage tasks.
OpenAI said those results came largely from experiments in which researchers explicitly instructed the model to evade monitoring. The company also reported that Astra was less likely overall than GPT-5.6 Sol to violate safety and security restrictions.
Still, the findings indicate that increased capability does not automatically make an AI system easier to understand or supervise.
OpenAI says it has placed monitoring on all externally deployed Astra interactions involving tools and continues to research oversight methods that do not rely solely on examining a model’s internal reasoning.
What Users Should Expect
Astra still may not appear in every ChatGPT account or every product surface at once. OpenAI has described the release as a phased rollout, and availability can differ among Plus, Pro, Business and Enterprise plans.
When access arrives, users can expect a model intended to handle longer, more complicated assignments with less step-by-step direction. Human review will remain important, particularly when AI is given access to accounts, private records, workplace systems or the ability to take actions online.
Provided by Dallas Express
