Gemini 4 Argon: Google Gates AI Bug-Hunter + Mistral Large 4

Samantha: Welcome to The State of Tech, The European Edition, Friday October nine, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Google locks down a powerful new AI model built to hunt software bugs, OpenAI launches persistent personal agents called Dots, Washington freezes green card applications for a long list of big tech employers, European champion Mistral previews its one-trillion-parameter Large 4, and to close out, two things you can actually use today: the open-source note app Zotero, and the anonymous messenger Keet. Let's start with Google.

Google unveils Gemini 4 Argon, a frontier AI model that hunts and fixes critical software bugs, with access deliberately limited.

Samantha: Google has introduced Gemini 4 Argon, and it is a notable moment because the company is not throwing it open to everyone. This is a frontier model built with one job in mind: cybersecurity. It can autonomously spot software vulnerabilities, verify them, and in many cases propose a working fix. Google says it has already found a critical flaw in healthcare software that earlier models missed. Access runs through a tightly controlled programme called Fairwind, open only to vetted cybersecurity experts.

Bob: The restraint is the real story. Google explicitly says it is holding back wider access because the same capability that patches a hole can also be used to find and exploit one. That is a shift in tone from the usual race-to-ship we have seen from the big AI labs. The plan is a gradual expansion, more safeguards, more risk assessments, then broader rollout. Think of it as a pharmaceutical trial model applied to AI, where capability alone is not enough to earn distribution.

Samantha: The healthcare example matters because that sector sits on ageing code and strict regulation. If a tool can quietly catch a critical bug in a hospital system before an attacker does, you shorten the window where patients are at risk. The flipside is that if the same tool leaks or gets replicated by less careful players, you hand attackers a bug-finding engine. That is why gated access, rather than a public API, is the central design choice here.

Bob: For European security teams and software vendors, the question is practical: will they be invited into Fairwind, and under what conditions? European firms often sit outside the first wave of US pilot programmes, which creates a dependency problem if critical infrastructure vendors have to wait months longer for the same defensive tooling. Regulators in Brussels will also want to see how Google proves that an autonomous bug-finding and bug-fixing agent stays inside its intended scope.

Samantha: The broader signal is that frontier AI is quietly moving from general assistants to specialised, high-stakes agents. A cybersecurity model with autonomy is not a chatbot; it is closer to a self-directed junior analyst with root access to understanding your code. Expect this pattern, narrow powerful tools behind heavy gates, to spread to other sensitive domains like finance, biosecurity and critical infrastructure. The era of "release first, govern later" is being tested in real time.

OpenAI launches Dots, persistent personal AI agents that keep working in the background across thousands of apps.

Bob: OpenAI has launched something called Dots, and the pitch is a step beyond ChatGPT as we know it. Each Dot is a personal AI agent assigned to a specific goal or responsibility. It runs in its own cloud environment, works in the background, and keeps going even when you close the app. OpenAI says Dots can plug into more than four thousand services, which means a single agent can gather information, update projects, and take actions across many tools without you prompting each step.

Samantha: The important shift is from reactive assistant to persistent worker. Today most people use AI by opening a window, typing something, and waiting. A Dot is more like a background colleague who checks your inbox, drafts replies, moves tasks across a project board, and keeps a long-running task alive for days. For routine admin, that is genuinely useful. For sensitive work, it raises an obvious question: how much authority are you comfortable handing to a process you cannot see?

Bob: Integration with four thousand apps is both the selling point and the risk. Every connection is a potential path for data to flow out, and the agent's decisions depend on how well permissions are scoped. If a Dot has access to your calendar, your drive, your bank integrations and your work chat, a single misconfigured instruction could reach further than you intended. Enterprises will want very clear audit logs before they roll this out to staff.

Samantha: For European users and companies, the GDPR implications are concrete. Persistent agents that read, store and act on personal data across multiple services need a lawful basis for each step, and the data must stay within agreed boundaries. Data protection authorities have already signalled interest in autonomous agents, and Dots fits precisely into that scope. Expect questions about where the agent's cloud environment physically runs, and how a user can see and delete what their Dot has learned.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Washington freezes green card applications from Microsoft, Infosys, TCS and others over foreign-worker allegations.

Bob: The US Department of Labor has suspended access to a key employment-based green card pathway for a long list of big tech employers. The names include Microsoft, Adobe, Cognizant, Infosys, Tata Consultancy Services, Wipro, HCL Technologies and Capgemini. The department says it will stop accepting new permanent labour certification applications from these firms and halt processing of applications already in the queue. The allegation driving the move: that these companies prioritised foreign workers over American job applicants.

Samantha: The direct impact is on the sponsorship route many H-1B visa holders use to move to permanent residency. For engineers already in the US on temporary visas, that pathway just closed at these employers, which creates real uncertainty about their long-term status. It also forces the companies to rethink how they staff American projects, because a frozen green card pipeline makes it harder to retain senior foreign engineers who were planning to stay.

Bob: There are two European angles worth noting. First, several of the named firms, including Capgemini and the Indian IT majors, are significant employers across Europe, so internal talent flows between US and European offices may shift. Second, if skilled engineers find the US route blocked, European tech hubs, Dublin, Amsterdam, Berlin, Paris, become more attractive destinations. We have seen this pattern before during past US visa squeezes, and recruiters here tend to move quickly when it happens.

Samantha: The broader point is that immigration policy is becoming an active instrument of tech industrial strategy. Freezing a specific group of employers, rather than reforming the system as a whole, sends a political message as much as an administrative one. For European policymakers trying to attract AI and cloud talent, this is an opening. Whether the continent can convert that opening depends on housing, salaries, and whether the EU blue card process can actually move faster than Washington's.

Mistral AI previews Large 4, a one-trillion-parameter European frontier model trained entirely in EU data centres.

Bob: European AI company Mistral has previewed its fourth-generation flagship, Mistral Large 4, internally nicknamed Le Chonk. The headline numbers: one trillion total parameters, with forty-nine billion active per token, and it is natively multimodal, meaning it handles text, images and other inputs in one model. Mistral trained it from scratch on three thousand eight hundred NVIDIA Grace Blackwell chips inside the company's European data centres. A guardrailed preview is live now through Mistral Studio, with full weights expected later this month.

Samantha: Two things stand out. One, the training was done entirely on European infrastructure, which matters politically as much as technically. European policymakers have been pushing for sovereign AI capacity, meaning models trained, hosted and governed inside the EU, and Mistral is now the clearest example of that in action. Two, the training data covers more than one hundred sixty languages, including every official EU language. That is a direct answer to the complaint that frontier models underperform in smaller European languages.

Bob: The release of the full weights is the detail to watch. If Mistral really does ship the model openly by the end of the month, European companies, research institutions and public bodies can run it on their own hardware, under their own rules, without sending data to a US provider. That is a very different offer from what OpenAI or Google provide. It also puts pressure on competitors to explain why their most capable models remain closed.

Samantha: The practical limits are worth naming. One trillion parameters is a lot to run; most organisations will not host this model themselves and will use the Mistral Studio API instead. And a guardrailed preview means capabilities and behaviour are still being tested. But even with those caveats, this is the strongest signal yet that Europe can produce a frontier model on its own soil, in its own languages, on its own chips, rather than only consume what the American labs decide to release.

Zotero collects, organises and cites your research sources in one free, open-source library you control.

Bob: To close out, two things you can actually use today, and the first is a free research tool called Zotero. It is an open-source reference manager that gathers articles, web pages, books and PDFs in one personal library on your own computer. You save a source with one click, and Zotero stores the details for you. No more folders full of files called final-final-version.

Samantha: What makes Zotero different is the browser connector. When you find an article or a web page, you click a button in Firefox, Chrome or Safari, and Zotero captures the title, author, date and often the PDF itself. Later you can add tags, group sources into collections, and attach your own notes to any item. Users describe it as particularly useful for theses, reports, journalism research and long-term reading projects. And the search covers the full text of your PDFs.

Bob: It runs on Windows, macOS and Linux, and the software itself is free. The learning curve is small: within about fifteen minutes you can install it, add the browser connector and save your first dozen sources. There are plugins for Word and LibreOffice that insert citations and build a bibliography in thousands of styles, from APA to Harvard. For anyone tired of formatting references by hand at midnight, that alone is worth the download.

Samantha: The honest caveats: Zotero gives you a modest amount of free cloud storage for syncing files, and if you want more you pay a small yearly fee, or you can use a WebDAV server instead. The interface is functional rather than beautiful, closer to a traditional desktop application than a slick web app. And it is built for sources and citations, so it is not a general notebook for everything in your life. But if you want an organised library that does not lock your work away, those trade-offs are easy to accept.

Bob: It is particularly popular with students, academics and anyone who writes with sources and wants a stable home for them over years. Reviewers often praise it precisely because it is community-driven and has stayed true to its purpose. You can export your whole library in open formats, and that is a quiet promise many commercial research tools simply cannot make.

Keet lets you video call, chat and share files directly with friends, with no account, no sign-up and no central server in the middle.

Samantha: And the second thing to try today is an app called Keet. It is a peer-to-peer video calling and messaging app from the team behind Holepunch. The idea is simple: you open it, you create a room, and you share a link or key with the person you want to talk to. There is no account to register and no server sitting in the middle of your conversation.

Bob: The way it works is unusual. Keet connects your device directly to the other person's device, so your calls and messages travel between you rather than through a company's data centre. The conversations are end-to-end encrypted, and because there is no central service collecting everything, there is no big database of who talked to whom. For anyone who has grown uneasy about handing every video call to a large platform, that is a refreshing change.

Samantha: It is free to download, and it runs on Windows, macOS, Linux, Android and iOS. Rooms can hold groups, and you can share files and photos in the same place as your chat. The catch is that it is a younger project than the big names. Fewer of your friends will already have it installed, and you will need to send them a room link and ask them to download the app first.

Bob: The honest trade-off is simple: more independence, less familiarity. For a quick call with a colleague, the usual tools are still the easiest option. But if you want to see what communication looks like when there is no company in the middle, Keet is worth half an hour of your afternoon. Set up a room, invite one friend, and notice how little it asks of you. It is a glimpse of how calls and chats can work when the design starts from the idea that your conversation belongs to you.

Bob: Today we covered: Google's gated Gemini 4 Argon for cybersecurity, OpenAI's persistent Dots agents, the US green card freeze hitting major tech employers, Mistral's one-trillion-parameter Large 4 trained in Europe, and two things to try yourself: the offline knowledge-base app Zotero, and the anonymous messenger Keet.

Samantha: Today we covered: Google's tightly gated Gemini 4 Argon for cybersecurity, OpenAI's new persistent agents called Dots, Washington freezing green card applications at major tech employers, Mistral's European-trained Large 4 frontier model, and two things to try yourself: the private notebook app Zotero, and the anonymous messenger Keet.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.