OpenAI launches cyber AI model, Analog Devices buys Alif

Samantha: Welcome to The State of Tech, The European Edition, Saturday November seven, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: OpenAI builds an AI model that hunts for software weaknesses, chip packaging becomes the new choke point of the AI boom, Analog Devices closes a billion-dollar deal for chips that run AI on the device itself, computing giant IBM unveils its most advanced quantum computer in Chicago, and two things to try yourself: Physicl Studio, a library of virtual worlds for training robots, and a quick security check of your own GitHub repositories. Let's start with OpenAI.

OpenAI launches a specialised AI model that finds, tests and helps fix software vulnerabilities.

Samantha: OpenAI has launched a new artificial intelligence model built specifically for cybersecurity. It's designed to find weaknesses in software, test whether an attacker could actually exploit them, and then help write the fix. That's the full cycle, from discovery to repair, in one tool. But you can't just sign up. OpenAI has rated the model Critical under its own risk framework, the Preparedness Framework, and for now access is limited to verified participants in a program called Daybreak Red.

Bob: The performance jump stands out. On a test called ExploitGym, which measures how well an AI can turn a known weakness into a working attack, the new model scored forty-eight and a half percent. OpenAI's previous cyber model scored thirty-two point one percent on the same test. That's roughly half as good again. And notice what that number measures: not spotting bugs, but exploiting them. That's the step defenders use to prove a problem is real, and the step attackers use to break in.

Samantha: That dual-use problem explains the tight gate. A tool that shows a company where its software breaks can, in the wrong hands, show an attacker the same thing. The Critical rating signals that OpenAI itself sees serious risk in this capability, and limiting it to vetted participants is the company's way of pointing it at defence first. The source also describes this as initial access, which suggests the circle could grow later. How well the gate holds then is an open question.

Bob: There's a second tension, and the reporting flags it directly. AI can now find problems faster than organisations may be able to repair them. Finding a bug is one thing. Fixing it means testing a patch, getting it approved and rolling it out across thousands of machines, sometimes in systems that can't simply be switched off. If discovery speeds up and repair doesn't, the pile of known but unfixed weaknesses grows. The model does help draft fixes, but someone still has to ship them.

Samantha: For Europe, this lands on two groups in particular. European cybersecurity firms could get far more powerful tools for auditing software. And operators of critical infrastructure, think energy networks, hospitals and transport, face the flip side: more weaknesses surfacing, faster, with pressure to patch on a shorter clock. The bottom line is that the bottleneck in security may be shifting from finding problems to fixing them. According to the reporting, that means faster repair processes, not just better scanners.

Advanced chip packaging has become the critical bottleneck in the global AI supply chain.

Bob: The biggest bottleneck in AI hardware right now may not be making chips at all. It's packaging them. Packaging is the step after a chip is made, where it's mounted, connected and sealed so it can go into a server. Advanced packaging goes further: it combines several chips, side by side or stacked, into one unit that works as a single, more powerful processor. Demand for AI processors is now outrunning the capacity to do that.

Samantha: And the money is following. Nvidia, AMD and TSMC are all investing heavily to expand their packaging capacity. Industry analysts expect capital spending on advanced packaging to reach nearly twenty-eight billion dollars, and they forecast that this year, for the first time, revenue from advanced packaging will overtake traditional packaging. For decades this was the unglamorous back end of chipmaking. Now a cutting-edge chip that can't be packaged is, in practice, a chip that can't be shipped.

Bob: The bigger story is a shift in how progress happens. For a long time, the main route to faster chips was shrinking transistors, the tiny switches inside them, so more fit on each piece of silicon. That still matters, but analysts describe a fundamental transformation: performance increasingly comes from integrating multiple chips. That changes where leverage sits in the supply chain. Packaging capacity now helps decide how many AI processors actually reach customers, not only the factories that make the chips.

Samantha: For Europe, the reporting points to supply chain resilience. European chipmakers and AI developers rely on this same step, so a shortage in packaging capacity can slow their access to AI hardware too. It also widens what competitiveness means in chips. Designing or manufacturing a processor is only part of the picture if the final assembly step is where capacity runs short. Key takeaway: the question for European industry is no longer just who makes the chips, but who puts them together.

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.

Analog Devices closes its 1.35 billion dollar purchase of edge-AI chipmaker Alif Semiconductor.

Bob: Analog Devices has closed its one point three five billion dollar purchase of Alif Semiconductor, a maker of chips that run artificial intelligence directly on devices. The deal was completed on November third. Alif makes low-power microcontrollers, small chips that control a single device, plus processors designed for AI at the so-called edge. The edge simply means the device itself: a factory sensor, a robot arm, a medical monitor, rather than a distant data centre.

Samantha: The logic is about filling a gap. Analog Devices already has businesses in sensing, signal processing, power and connectivity. In plain terms: measuring things in the physical world, cleaning up those measurements, powering the hardware and getting data off the device. What Alif adds is the ability to run AI on that same device. So the pitch is a fuller package for customers who want a machine that senses something and makes a smart decision locally.

Bob: The company names four target markets: industrial systems, data centre infrastructure, robotics and digital health. Health is easy to picture. A bedside monitor that interprets data on the spot can react faster and depends less on a network connection. In factories, local processing lets a machine flag a problem as it happens. What this deal shows is that the AI chip race isn't only about giant processors in data centres. There's a parallel race for small, efficient chips inside everyday machines.

Samantha: For Europe, the reporting frames this as a shift in the competitive landscape for semiconductor firms and AI hardware developers. European companies in the same space now face a rival that offers sensing, power, connectivity and AI processing under one roof. For European device makers buying these components, that could mean simpler one-stop shopping, but potentially one fewer independent supplier to choose from. This deal puts a concrete price tag on on-device AI capability.

IBM unveils its most advanced quantum computer at the University of Chicago's Hyde Park Labs.

Bob: Computing giant Eye-Bee-Em has unveiled its most advanced quantum computer, and it's in Chicago. Together with the University of Chicago and the State of Illinois, the company installed the system at the university's Hyde Park Labs. It's the first Eye-Bee-Em quantum computer in Illinois. At its heart is the company's Nighthawk processor, the chip that performs the quantum calculations.

Samantha: The machine isn't meant to be a showpiece. It's intended to anchor a National Quantum Algorithm Center. Algorithms are the step-by-step recipes a computer follows, and quantum computers need entirely different recipes from ordinary ones. Writing them is one of the big unsolved jobs in the field. Powerful hardware is one thing; knowing which useful problems to run on it is another.

Bob: The partners name concrete targets: drug discovery and more efficient power grids. The plan is to bring universities, national laboratories, startups and industry together around the same machine, so ideas can move from research towards real applications faster. That's the idea behind building a shared hub around the hardware.

Samantha: Expectations deserve some precision here. The announcement is about accelerating the development of algorithms for real-world applications. It doesn't claim those applications are ready today. This is research infrastructure: a place where the hard work of making quantum computers useful can happen. And unlike much of today's news, there's no AI hook. It's an entirely different kind of computing.

Bob: For Europe, the deployment is a marker in the global race for quantum advantage, the point where quantum machines beat ordinary computers on useful tasks. European research groups have a stake in that race, and collaborations are one route in. The set-up in Chicago, hardware anchoring a hub of universities, labs and startups, is also a model European regions can measure their own quantum efforts against.

Physicl Studio offers ten thousand virtual environments for training robots on synthetic data.

Samantha: To close out, two things you can actually look at today. First: Physicl Studio. Physicl is a startup working on physical AI, the software that helps robots understand and interact with the real world. It just raised ten million dollars in seed funding from Inovia Capital and launched Studio, a platform of realistic 3D simulations for training robots. Think of it as a vast library of virtual rooms and objects where a robot can practise before it touches the real thing.

Bob: The numbers give a sense of scale: ten thousand ready-made environments and a hundred thousand digital objects to fill them. The goal is synthetic training data, meaning realistic practice material generated by a computer instead of filmed in the real world. That matters most for edge cases, the rare situations that are hard to capture on camera. Think of a box sliding off a shelf at the wrong moment. Hard to stage over and over in reality, simple to repeat in a simulation.

Samantha: The company's argument is that training robots on real-world video is expensive and often ineffective, and that simulation is a more scalable and accurate route. The practical step today is to look up Physicl Studio online and see what these virtual training worlds look like. One clear caveat: the announcement doesn't spell out pricing or whether individuals can sign up. It's built for robotics teams, so treat it as a window into how robots are taught, not a consumer app.

Bob: For people working at a European robotics company, a university lab or a startup building autonomous machines, there's a more direct use: exploring whether simulated training could replace some costly real-world recording. The reporting points to more efficient and safer training methods for autonomous systems. For everyone else, it's a concrete look at one way robots learn their skills: inside virtual worlds, before they're let loose in ours.

A quick check of your public GitHub repositories protects against AI tools hunting exposed keys.

Samantha: The second thing costs nothing but a few minutes: a security check of your own GitHub repositories. GitHub is where many developers, students and hobbyists store and share code. Microsoft Threat Intelligence reports that hackers are using automated AI tools, including one called Artex, to scan public GitHub repositories for exposed passwords and API keys. An API key is a digital pass that lets one program log into a service, like a cloud account, without anyone typing a password.

Bob: What happens next is the serious part. According to Microsoft's researchers, stolen keys are used in follow-up attacks on the organisations they belong to. In one case, a single internet address used stolen application credentials to sign into several different company cloud environments. From there, data was pulled out of Microsoft's Azure cloud, including from vaults meant to store secrets, and the attackers also abused Azure's AI services. Because the scanning is automated, nobody has to stumble across a leaked key by chance.

Samantha: So here's the check. Open your GitHub account and go through every repository set to public. Ask whether it should be public at all; if not, switch it to private. Then look through configuration files and code for anything resembling a password, token or key. If you find one, deleting the line isn't enough, because older versions of files stay in the project's history. The safe move is to cancel that key at the service that issued it and create a new one.

Bob: A few limits to keep in mind. This check only covers your own public code; it does nothing about keys pasted into chats, documents or other services. Making a repository private stops new visitors, but it doesn't undo what may already have been copied, which is why replacing the key matters. It works in any browser and needs no extra software. The reporting stresses that European companies using GitHub and cloud services face this risk too, so the same routine applies to code you manage at work.

Samantha: Today we covered: OpenAI's new model for finding and exploiting software weaknesses, chip packaging as the AI bottleneck, Analog Devices buying Alif Semiconductor, Eye-Bee-Em's new quantum computer in Chicago, and two things to try yourself: Physicl Studio and a security check of your own GitHub repositories.

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.