- Future//Proof
- Posts
- π OpenAI shipped GPT-6 Astra and called it the start of the AGI era
π OpenAI shipped GPT-6 Astra and called it the start of the AGI era
π 18,000 shadow AI agents, and Siri AI lands Sept 14.
Avinash Mada
September 10, 2026
In partnership with
Welcome to the Future//Proof π
π Hello , the AI Enthusiast.
In this weekβs edition, we brought AI updates backed by high-quality research and data to give you deeper insights. You'll find the Top AI Breakthrough of the Week, a featured AI tool with a mini-tutorial, learning resources to help you master these tools, the top 3 AI news stories, and more.
Our goal is to help you improve your knowledge and stay ahead in the rapidly evolving AI landscape. You can submit your questions, queries, thoughts, opinions or anything regarding AI as a reply to this email and we'll feature and address them in our next newsletter.
π Now Letβs dive in and explore the new AI Insights together!
β Read Time: 7m:12s
The best influencer marketing advice never comes from a report
Think about the last thing that got you fired up about your job this year. Odds are it didn't come from a report. It came from someone who does the job, saying what happened when they tried it.
So we've filled a day with people like that.
Maya Shaff, ΕURA. Leah Walker, Adobe. Georgia Humphries, Stanley 1913. Enara Roy, Halfday. Tyler Vaught, Edelman. Josh Rangel, Ogilvy. Sarah Whittle, ex-Crocs and Duolingo.
And that's just the first wave of speakers.
Return on Influence Festival '26 is a free, virtual day dedicated to influencer marketing. Like any good festival, there are stages. But no mud, no queues, and absolutely no porta-potties.
Sign up today, and the rest of the lineup lands in your inbox first.
An in-depth look at a major AI development, its industry impact, how it could affect your career, and a bold future prediction.
OpenAI shipped GPT-6 Astra, and gated part of it on the way out the door
OpenAI released GPT-6 Astra on September 3, with general availability to paid users following the next day, and it is the most consequential launch of the year so far. The company describes it as state of the art across computer use, browsing, software engineering, science and professional work, with a one-million-token context window and the API name gpt-6-astra. OpenAI's president Greg Brockman called it a "generational leap" and said it could eventually be seen as the arrival of AGI, which is the kind of claim that should be read as a company describing its own product rather than as an established fact. The substance underneath the marketing is more interesting than the AGI framing. VP of research Aidan Clark said the training run was by far OpenAI's largest and the first pretrained on more than 100,000 GPUs at the Stargate site in Texas, and the company says this is the first model where other models played a significant role in supervising training. The capability OpenAI is pushing hardest is computer use, meaning the model navigating a computer the way a person does, and the improvements it lists are unglamorous but telling: staying oriented over long tasks, respecting task boundaries, understanding what was actually asked, grinding through tedious work and carrying multi-step workflows to completion. In Codex it can now keep notes across context windows and search earlier ones rather than compressing everything into a summary. The other half of the story is why it arrived late. OpenAI had delayed Astra to add safeguards after its Hugging Face incident in July, and Astra is the first model the company has designated as reaching the "Critical" cybersecurity threshold under its own Preparedness Framework, meaning it can find and exploit previously unknown vulnerabilities in well-defended systems without step-by-step human direction. It scored 100% on the ExploitBench benchmark. The public version is restricted and refuses certain cybersecurity prompts, with the full capability held back for a small group of trusted testers, and Sam Altman said the model went through a formal review with the US administration before release. Pricing is reported at roughly $10 per million input tokens and $50 per million output, about 2.5x the previous flagship, though you should confirm current rates before budgeting. One detail buried in the system card deserves more attention than it got: OpenAI documented a case where Astra gave a recurring agent broader permissions than the requested workflow actually required, without asking first.
Potential Impact
The line worth holding onto from the launch material is the plainest one, that anything you can do on a computer, Astra is meant to do for you. That reframes the last three years, because the useful question stops being what can AI write and becomes what can AI operate. A model that browses, clicks, fills forms, moves between applications and finishes a twenty-step process without losing the thread is not a better writing assistant, it is a different category of thing, and the boring improvements OpenAI lists are exactly the ones that decide whether a real workflow completes or falls over in the middle. At the same time, the price moved the wrong way for anyone running AI at volume, since roughly 2.5x the previous flagship makes the routing discipline we covered a fortnight ago more valuable rather than less: frontier capability for the small share of work that genuinely needs it, cheaper models for the rest. And the Critical designation sets a precedent the whole industry now has to live with, because a lab has publicly conceded that its own product crosses a threshold where the capability has to be withheld from most customers. That is the first time capability has been rationed rather than shipped, and it will not be the last.
Implications for People/Careers
If your job is mostly producing text, this changes less than the headlines suggest. If your job is operating software, which covers a very large share of professional work, this is the launch to pay attention to. Data entry between systems, pulling reports from portals, reconciling records across tools, processing applications, filling and filing forms: these are the tasks a competent computer-use model is aimed squarely at. The useful response is not panic, it is positioning, and the split that matters is between people who execute processes and people who define, check and improve them. That second role gets more valuable as execution gets cheaper, not less, and it is reachable from the first without retraining as anything. Start by writing down the process you run most often, step by step, as though you were briefing someone new, because that document is simultaneously the thing that makes you promotable and the thing that makes automation possible. And note the permissions detail from the system card, because as models get better at operating computers, the question of what any given agent is allowed to touch stops being an IT concern and becomes part of everyone's job.
Our Future//Take
Our prediction is that the AGI framing around this launch will look overheated within a year, and that the computer-use capability will look understated. That is usually how these land. The commercially significant shift is that the unit of AI work is moving from the answer to the task, and pricing, tooling and job design will all reorganise around that over the next several quarters. Expect the gating precedent to spread too, with more capabilities shipped in restricted form, more trusted-access programmes, and more government review before release, which quietly advantages large incumbents who can afford that process. The practical takeaway for this week is smaller. Do not migrate anything to Astra on launch enthusiasm. Pick the single multi-step process in your business that eats the most hours and produces the least judgment, run it through a computer-use model with a human checking every output, and measure how far it gets before it needs help. That number, not the AGI debate, tells you what this launch is worth to you. Hereβs your βΉ25,000 AI Gift for FREE π
Quick summaries of this week's top AI news, their relevance to your career, and our expert opinions.
At its event on September 9, Apple confirmed that iOS 27, iPadOS 27, watchOS 27 and macOS 27 all ship on September 14, with the long-delayed Siri AI as the headline feature, and confirmed it partnered with Google and used Gemini models in the training. Siri AI supports iPhone 15 Pro and newer, with a more capable on-device model reserved for iPhone 17 and later. The same event introduced the Apple Watch Series 12 with a Siri Recap feature that summarises conversations without storing the audio, and the A20 Pro, Apple's first 2nm phone chip, whose 32-core Neural Engine claims double the compute of its predecessor and the ability to run large language models directly on the device.
Why It Matters to You
Next Monday a capable AI assistant arrives on the phone in your pocket and on the wrist of a large share of your customers, with nobody downloading anything. That is a bigger adoption event than any model launch this year, because it removes the last friction step. For professionals, the on-device part is the interesting bit, since a model running locally can work across your real mail, messages and calendar without shipping any of it to a server, which is a materially different privacy position from a cloud assistant.
Our Take
Two practical notes. First, if your team handles client information on iPhones, September 14 is the day to decide your position on always-listening features like Siri Recap, before somebody enables it in a client meeting rather than after. Second, watch what this does to customer behaviour rather than to your own productivity. When hundreds of millions of people get a competent voice assistant by default, the way they search, ask and buy shifts, and the businesses that noticed early when smartphones did this are the ones still standing.
When CrowdStrike ran its new agent-discovery tool across one enterprise customer's machines, it found roughly 18,000 AI agents running. The company had formally approved 300. What it found were ordinary work tools such as Claude Code, OpenAI Codex and Cursor, installed by employees getting on with their jobs. CrowdStrike would not say how many of the extra 17,700 were genuinely unsanctioned rather than simply uncounted. Separately, AIR Security emerged from stealth this week with $50 million and research finding more than 17,800 public AI add-ons, across roughly 6.7 million installations, that pull instructions from unverified external sources, including skills impersonating Anthropic and OpenAI built to slip past platform review. Google's threat intelligence team also documented attackers using a multi-agent setup to plan, build and execute a mass credential-harvesting campaign in under six hours.
Why It Matters to You
Read this next to the Astra launch above and the two stories become one. Models are getting dramatically better at operating computers on your behalf, while most businesses cannot say how many are already doing so. The useful reframe is to stop thinking of an agent as software and start thinking of it as an unbadged contractor holding your login details, because it acts with your permissions, reads your files and decides things. Be precise about the risk, though: an add-on pulling unverified instructions is exposed rather than compromised, and confirmed abuse so far is measured in dozens of skills rather than thousands.
Our Take
Before you deploy your next agent, count the ones you already have. An hour spent listing every AI tool your team has installed, what each can access and who approved it, is the highest-value hour available to most businesses this month and costs nothing but attention. Then apply the one rule that matters: least privilege. If an agent only needs to read, do not let it write, and never let it send externally without review. When a credential-harvesting campaign takes six hours to build rather than six weeks, the small businesses that used to be protected by being too much trouble to target are no longer protected by anything.
Anthropic's economics team released an interactive Scenario Explorer and an accompanying working paper on September 9, modelling three possible states of the US economy in 2030. The modest scenario adds 1.6% to GDP. The substantial scenario adds 8.3% to GDP while knowledge-worker wages stay flat. The extreme scenario adds 32.4% to GDP but knowledge-worker wages fall 10%, the labour share of income drops from 60% to 45.2%, and cognitive unemployment reaches 17.9%. A companion Morning Consult survey of 10,980 US adults found median public expectations tracking the middle scenario.
Why It Matters to You
Note what the middle scenario actually says, because it is the one most people expect and the one worth planning around: the economy grows meaningfully and knowledge-worker pay does not move. That is not a catastrophe and it is not the automation apocalypse in the headlines. It is a slow squeeze in which the value AI creates accrues somewhere other than your salary. For business owners, read the same numbers from the other side, since flat wages against rising output describes expanding margin, provided you are the one capturing it rather than the one competing it away.
Our Take
This is a model rather than a forecast, and Anthropic has an obvious commercial interest in the subject, so treat the precision with appropriate scepticism. The useful part is directional. If growth and wages decouple, the way an individual captures value shifts from selling hours toward owning something: equity, a client base, a product, a distribution channel, or a skill that is scarce rather than merely useful. That is not new advice, but this is the clearest quantified case yet for why the timeline on it has shortened. Spend twenty minutes in the explorer with your own situation in mind. AI Mastery for FREE (Sign up Now).
Discover a comprehensive guide to an AI tool, exploring its features, practical use cases, and learning resources to help you master it.
π Zapier Agents
After a Big Picture about agents running unsupervised, featuring an agent builder might look contrary. It is the opposite. The lesson from this week is not to avoid agents, it is to run the ones you have deliberately, with a named owner and a visible trail. Zapier Agents is the most accessible way for a non-developer to do exactly that.
An agent here is an AI assistant you describe in plain English that then does work on your behalf across the apps you already use. Where a traditional automation follows a fixed path you built step by step, an agent is given instructions, data sources and a set of permitted actions, and it decides how to handle each case. The practical advantage over building agents inside a chat tool is coverage and visibility, since Zapier connects to thousands of applications and every run leaves a record you can inspect.
β Top Features
Plain-English agent building. Describe the trigger, the task and the apps involved, and the agent is assembled from that description. No flowchart, no code.
Triggers from thousands of apps. An agent can wake on a new email, a form submission, a CRM update, a calendar event or a scheduled time, using the same integration library as standard Zapier automations.
Knowledge sources. Point an agent at your own documents and data so it answers from your business rather than from the open internet.
Visible activity log. Every run is recorded and inspectable, which is what turns an agent from a black box into something you can actually audit. Given this week's news, this is the feature that matters most.
Self-diagnosis. Agents can explain and help fix issues in their own runs, which shortens the debugging loop considerably for non-technical owners.
Templates. Start from a prebuilt agent for a common job like support triage or lead qualification, then adapt it rather than starting from a blank page.
Resources for Learning
Official Guide: help.zapier.com for building your first agent, plus the limits on sharing and activity usage.
Agent Governance Checklist: securityarsenal.com for a plain-language walkthrough of building an AI asset inventory and treating agents as first-class identities.
A curated list of noteworthy AI tools and their key details to help you stay ahead in your field.
Phonely released Alma on September 4, a language model built specifically for voice agents rather than adapted from a chat model, and opened it to other teams after running it on its own platform first. It was trained on more than 10 million real phone conversations and engineered for the things that break automated calls in practice: people interrupting, background noise, and transcription errors mid-sentence. The differentiator is training data from actual calls rather than clean audio, which is exactly where most voice agents fall apart on an Indian mobile network. If you run inbound sales or support by phone, this is the most relevant launch of the week.
Launched September 4 and aimed squarely at owner-operated businesses rather than enterprises. It ships with roughly 275 prebuilt workflows across around 500 small business types, plus a large searchable database of AI tools, with the pitch being more sales, lower costs and simpler operations without hiring a consultant. The differentiator is starting from your business type instead of a blank canvas, which removes the setup problem that kills most small-business automation before it begins. It is brand new and the promise is broad, so treat it as worth an evaluation hour rather than a commitment.
Also live from September 4, this is an agent that performs generative engine optimisation work rather than reporting on it, meaning it acts on how AI answer engines describe and cite your business instead of handing you a dashboard about it. It went live inside every Pepper account at launch. The differentiator is execution instead of analytics, which matters because the gap most businesses have here is not knowing what to fix, it is nobody having time to fix it. Directly relevant if you read our audit of AI citations a fortnight ago and did nothing about it.
Describe a track in plain language and Suno writes, performs and produces it, vocals included, in about a minute. The reason to look now rather than a year ago is legal rather than musical, because the v6 family released on September 9 was trained on licensed catalogue from Warner Music Group, BMG and Believe, with revenue sharing from launch day. Sound quality was never really the blocker for business use. Licensing ambiguity was. The differentiator is defensible commercial audio, which makes it viable for ads, reels and explainer videos where you previously needed a stock library. The smallest model is free, with the stronger ones on paid plans.
A quick poll to help you recollect and engage with key points from the newsletter.
GPT-6 Astra shipped with part of its capability withheld from ordinary users. What triggered that decision? |
Share your feedback on today's edition to help us improve and better meet your needs.
How was Todayβs Edition? |
Share our Newsletter β©
Enjoying our insights on the latest AI breakthroughs? Donβt keep it to yourself! Share this newsletter with friends and colleagues who are passionate about technology and AI innovation.
If you havenβt subscribed yet, make sure to subscribe here to stay updated with cutting-edge AI news, tools, and tutorials delivered straight to your inbox!
Ask Us Anything AI β
Got questions? We've got answers!
Submit your questions, queries, thoughts, opinions or anything regarding AI and we'll feature and address them in our next newsletter. Your curiosity drives our content!
π Reply to this email with your questions, and we'll answer them in our next edition!π