OpenAI Released a Model So Powerful Their Own Safety Team Called It a Critical Risk

On September 3, 2026, OpenAI launched GPT-6 Astra, an AI model that autonomously operates software to complete tasks, marking what the company calls the start of the AGI era. OpenAI's own safety team flagged the model as a critical risk due to its 'significant jump in cyber capabilities,' and the model was trained with supervision from other AI models, enabling faster self-improvement. With 100 million users gaining access, Astra is already deployed and changing how people work.

On September 3, 2026, OpenAI launched GPT-6 Astra and declared the AGI era. Here is what the headlines didn’t explain — and why it matters for everyone who works for a living. On September 3, 2026, Everything Changed. Sarah sat at her desk, sipping her morning coffee. It was a routine she followed each day without thinking — check emails, plan her tasks, and start her work. But on September 3, 2026, the world shifted under her feet in a way she had not yet realised. This Is Not What You Think It Is Think about every AI you have used before. You opened it. You typed something. It responded. You took that response and went and did something with it. You were always the operator. The AI was always the tool. Astra reverses that relationship entirely. You give it a goal. It opens the browser. It logs into the apps. It reads the files. It completes the tasks. It sends the results. You come back to review. You are no longer the operator. You are the approver. That is not an upgrade. That is a different job description. Every major technology shift in history has felt incremental from the inside and enormous in hindsight. The spreadsheet didn’t feel like it would eliminate entire accounting departments when it launched. The email didn’t feel like it would kill the memo. What you are reading about right now will feel the same way in ten years. The question is whether you understand it now or later. What Astra Actually Does Meet Marcus, a project manager at a medium-sized firm. His Monday mornings were usually chaotic. A board report due, meetings to schedule, competitor analysis to draft, and a presentation to build. Previously, Marcus juggled different apps and tools — Excel for reports, Outlook for scheduling, Google Docs for analysis, and PowerPoint for presentations. Each task required manual effort and time. With Astra, Marcus’s role transformed. He sets the goals: “Create a board report, schedule these meetings, complete competitor analysis, and prepare a presentation.” Astra dives into the apps, completes each task, and delivers the finished work. Marcus reviews the results. His hands-on workload shrinks drastically. Astra doesn’t just assist; it runs day-to-day operations autonomously. What Changed: Before Astra vs After Astra BEFORE: You → Ask AI → Get answer → You do the work → You use the tools → You send the result AFTER: You → Give Astra the goal → Astra uses the tools → Astra does the work → Astra sends the result → You review Welcome to the AGI Era Artificial General Intelligence AGI is the vision of AI capable of performing any economically valuable work as well as or better than a human. Think of a calculator — excellent at math but nothing else. ChatGPT excels at writing, but you still operate it. Astra is different — it autonomously uses real software to complete real tasks. At the end of the briefing, Greg Brockman said, “Welcome to the AGI era.” Those words could be the most significant in tech for 2026, or the most overhyped. You decide. Here is the detail that should make you pause. OpenAI said Astra is the first model where other AI models played a significant role in supervising its own training. That means the next version of Astra will be partially designed by Astra. The one after that even more so. The pace of improvement is no longer limited by how fast human researchers can work. It is limited by how fast the AI can improve itself. Whether Brockman’s “Welcome to the AGI era” is the most important sentence said in technology in 2026 or the most overhyped — you will not know for certain for several years. What you can know right now is this: OpenAI has 100 million users. They all woke up on September 4th with access to a tool that operates computers independently. That is not a demo. That is already deployed. Already in use. Already changing how people work. The Part Nobody Is Talking About OpenAI described Astra as a “significant jump in cyber capabilities,” meeting the “Critical threshold in cybersecurity” per their Preparedness Framework. They delayed its launch to install more safeguards, locking the most dangerous features behind a program called Daybreak. The public version is limited intentionally. Consider that. The creators were cautious enough to restrict it before release. Coincidentally, the same week Astra launched, Uber cut 3,300 jobs, Tesla’s driverless Cybercabs hit Austin streets, and Congress started inventorying rogue AI agents. These are not isolated events. They are interconnected pieces of a larger narrative. The unspoken question: If Astra can autonomously execute complex professional tasks, who decides which tasks it should perform? Who takes responsibility when it errs? When a company labels its product a critical risk, it’s not a marketing ploy. It’s an admission of power they are still grappling with. That merits more than passing interest. It demands deep reflection on our future trajectory. What This Means For Real People — Not Industries Priya, a marketing analyst, spends 6 hours every Monday pulling data from platforms, drafting reports, creating decks, and emailing them. With Astra, she gives a single instruction on Friday afternoon. By Monday morning, everything is ready for her review. Six hours of work compressed into six minutes of oversight. James, a software engineer, dedicates 30% of his time to code reviews, documentation, and mentoring junior developers. Astra doesn’t replace him. Instead, it manages those tasks alongside him, allowing James to focus on strategic architectural decisions. His productivity doubles, and his stress halves. Anika is a lawyer at a mid-size firm. Every morning she reviews contracts — checking clauses, flagging risks, comparing terms against precedents. It takes three hours before she can start billable client work. With Astra, she describes what she needs reviewed and the risk categories to flag. Astra reads the contracts, cross-references precedents, and produces a structured risk report with every clause annotated. Anika spends twenty minutes reviewing the report instead of three hours producing it. Her firm bills the same. Her morning is different forever. Sarah heads a team of eight content writers. They research, draft, edit, format, and publish. Astra can do all these tasks autonomously. Her company hasn’t announced changes yet, but the economic logic is clear. The question isn’t if this will affect Sarah’s team, but when — and whether Sarah will be proactive in shaping these changes or passive in reacting to them. What You Should Do Right Now If you are a knowledge worker of any kind : Try Astra this week. Not to replace your work — to understand what it can and cannot do. The people who understand its limits will be more valuable than those who either fear it completely or trust it blindly. If you are a manager or team leader : Have the conversation now. Do not wait for HR or leadership to tell you what to do. Understand which parts of your team’s work Astra could do, which parts it cannot, and what that means for how your team spends their time. The teams that figure this out first will outperform those that react to it later. If you are a business owner or executive : The question is not “should we use this?” The question is “what do we want humans on our team doing that AI should not do?” That is a values question as much as an efficiency question. Answer it before someone answers it for you. One thing applies to everyone : do not confuse speed with judgment. Astra can complete work faster than any human. It cannot yet decide what work matters, why it matters, or whether the result is right for the specific human reading it. That gap is where your value lives right now. Protect it. What Comes Next OpenAI built Astra on their largest training run ever, utilising over 100,000 GPUs. It’s also the first model where other AI models supervised its training. This self-referential training means Astra’s pace of improvement is no longer bounded by human intervention. The implications are profound. In the coming year, expect every major software company to integrate Astra-like capabilities into their offerings. Microsoft has Copilot, Google has Gemini, Salesforce has Einstein. The race to embed AI operators into workflows is already on. By year-end, the question won’t be whether your tools can do this, but whether you know how to direct them. The companies that move first will not just gain efficiency. They will gain a compounding advantage that is very hard to close. A team of ten people using Astra effectively can produce what previously required thirty. The gap between that team and a competitor who has not adopted it will widen every month. That is not a prediction. That is already visible in the companies using early access. The window to move early is measured in months, not years. The most valuable skill will not be using AI, but setting AI goals that achieve desired results. It’s harder than it seems. The gap between those who master this and those who don’t will be significant. The Bottom Line Four days ago, OpenAI declared, “Welcome to the AGI era.” You might have noticed the headline and scrolled past it. Here’s what you missed: This isn’t just a smarter autocomplete. This AI opens your browser, reads your files, uses your apps, completes your work, and closes the browser when done. Its launch was delayed because OpenAI’s safety team classified parts as a critical risk. Companies don’t say that lightly. The jigsaw puzzle analogy remains true, but stakes are higher now. For years, AI had pieces but lacked the framework to act. Astra has the framework, the pieces, and the hands to complete the puzzle without you. What you do with that information is the only question that matters now. Maduri Kolanu is a Senior Technical Lead specialising in AI-powered commerce experiences. She builds production-grade AI systems at the intersection of machine learning, customer behaviour, and enterprise commerce platforms. Follow for weekly insights on what is coming before it hits the mainstream. OpenAI Released a Model So Powerful Their Own Safety Team Called It a Critical Risk https://pub.towardsai.net/openai-released-a-model-so-powerful-their-own-safety-team-called-it-a-critical-risk-4a4e023513ac was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.