Last Week in Tech - Nvidia Buys Hugging Face, GPT-6 Crosses the Cyber Line, and Seven Exploited Flaws in a Day
20 minAudio
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Last Week in Tech for Monday 7 September 2026, covering 30 August to 6 September. Nvidia to acquire Hugging Face for US$12.93bn. Confirmed by Nvidia on 3 September, expected to close in the first half of 2027 subject to regulatory approval. Nvidia states the platform serves 18 million+ developers, 3 million+ models and 200,000+ companies, and has committed that Nvidia compute will not be required to build on or deploy through it. OpenAI ships GPT-6 "Astra" (3 September). OpenAI says it is the first model to meet the Critical cybersecurity threshold under its own Preparedness Framework - 100% on ExploitBench versus 78.5% for the prior model, and 39% versus 5.5% on a fresh-vulnerability variant. OpenAI says the model found and disclosed two previously unknown zero-days during evaluation. The shipped model refuses advanced offensive tasks. Separately, OpenAI committed US$1bn in subsidised defensive AI access for critical infrastructure, government, nonprofits and open-source maintainers; eligibility, subsidy level and post-subsidy pricing are not yet disclosed. Broadcom Q3 FY2026 (2 September). Revenue US$29.59bn, up 86% year on year; free cash flow US$13.67bn. CEO Hock Tan, quoted in the release, put AI semiconductor revenue at US$16.7bn, up 221% year on year, with Q4 expected at US$21.7bn. Longer-range AI revenue projections circulating this week came from earnings-call commentary, not the written release. CISA adds seven exploited flaws in one day (2 September). Four sit in AI and DevOps infrastructure - Kestra, JFrog Artifactory, Berri LiteLLM and Starlette. Microsoft reports attackers harvesting LiteLLM database tables for provider API keys and proxy-issued virtual keys, and advises monitoring AI workloads according to their control-plane role. Also on the list: Sangoma Switchvox (CVSS v4.0 9.3), a VoIP platform sold to SMBs, patched 14 July, first seen exploited 30 August. Fixed version 8.4.0.2. Elementor Pro under mass exploitation. Critical arbitrary file upload patched 19 August in version 4.2.2. Wordfence has blocked over 190,000 exploit attempts, and roughly two-thirds of Elementor's 10 million installations were still vulnerable as at 4 September. Indicator of compromise: any .php file under /wp-content/uploads/elementor/forms/. Chrome's sixth zero-day of 2026. V8 type confusion, fix shipped 3 September, added to CISA KEV 4 September. Fixed builds 152.0.7977.82/.83. Relaunch Chrome. Coming up: Wednesday, the PaperCut zero-day chain. Friday, 153 million driver's licence scans on the dark web. Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: NVIDIA newsroom; OpenAI; SecurityWeek; Broadcom investor relations; CISA KEV catalogue; Microsoft Security Blog; The Hacker News; Wordfence; BleepingComputer.
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Have you ever been sitting at your desk and you just feel this, I don't know, this profound sense of whiplash? Oh, absolutely. All the time. Right. Like you're reading an article in one browser tab about some trillion dollar leap in artificial intelligence. Yeah. You know, a new model that feels like literal science fiction. Yeah. Something that's just blowing every benchmark out of the water. Exactly. And then you click over to another tab and there is a notification telling you that you forgot to update your basic web browser. It's just so frustrating. It is. It's this bizarre
cognitive dissonance where, you know, we are living in the future, but we are still chipping over the digital extension cords of the present. That is well, I mean, it's really the defining tension of our current technological era. We are building these towering, incredibly sophisticated skyscrapers of machine intelligence. But we are essentially resting them on a foundation of. I was feeling this exact whiplash yesterday when I was reading through a tech dispatch from early
September 2026. The whole document captures this tension so perfectly. I knew we had to make it the focus of today's depth dive. That's a great source to dig into. Yeah, we are looking at a single week in tech that was split like cleanly down the middle. In the first half, we have the massive money and models fueling the AI boom. And in the second half, we have the immediate. real -world cyber vulnerabilities that are falling out of it landing right on your desk. Okay, let's unpack this. Because trying to navigate a landscape
moving this fast feels, honestly, almost impossible. It really does. But I think the trick here is to follow the money first. Okay, follow the money. Yeah, because to understand the daily security headaches you and your IT teams are dealing with, you really have to understand the macro. We need to trace the billions of dollars currently building the physical infrastructure of AI. Then we can look at the terrifying new capabilities of the frontier models that this money is actually producing. And finally, connect all of that high -level architecture down to the vulnerabilities actively
being exploited today. It's a single continuous loop. I love that approach. So before we even look at what AI is doing to cybersecurity, we really have to wrap our heads around the sheer scale of the infrastructure being built. Because the numbers are huge. The numbers in the source are just staggering. Let's start with NVIDIA. They are acquiring Hugging Face for $12 .3 billion. That is a massive acquisition. It is. And if you aren't like a machine learning engineer, that might just sound like another tech acquisition.
But Hugging Face is essentially the public square of open AI models. Yeah, it's completely foundational to the community. Right. It's a platform hosting 18 million developers and over 3 million models. And NVIDIA's CEO, Jensen Huang, made a very public point of... promising that Hugging Face will remain completely neutral, you know, multi -cloud and open -weight. Which is a big promise. It is. And if I understand open -weight correctly, that essentially means the core architecture of these models is out there for developers to download, run, and, you know, tinker with on
their own hardware rather than being locked behind some corporate API. That is the crucial distinction, yeah. Open weight models are really the democratic counterweight to the massive closed models run by companies like OpenAI or Google. Right, because you can actually see what's under the hood. Exactly. And NVIDIA is fully aware that Hugging Face's entire value proposition relies on this neutrality. If developers feel like the platform is just becoming, I don't know, a funnel to force them into buying NVIDIA hardware, they will just migrate
elsewhere. I hear that, but I have to push back a bit on the idea that things won't change. Oh, I'm sure they will, in some ways. Right, because if Hugging Face is the neutral Switzerland of the open -source AI world, NVIDIA just bought Switzerland for almost $13 billion. Do we really expect them not to plant their flag on the banks? Well, that's the billion -dollar question, isn't it? Literally. I mean, 18 months from now, when the regulators finally sign off, the licensing terms could easily shift. And the source makes a fantastic practical point here for anyone running a business. If your company relies on an AI model,
you need to write down exactly where it came from today. Yeah, it is the shadow IT problem, but supersized for the AI era. Oh, that's a good way to put it. Right, because 10 years ago, the problem was employees spinning up rogue Dropbox accounts. Today, it is a contractor who built your customer service chatbot using some open weight model pulled from Hugging Face. And nobody knows which one it is. Exactly. Nobody in your company actually documented which version it is or what license it operates under. If you
have a dependency you cannot name, you have a vulnerability you cannot manage. And it is not just software dependencies we have to worry about, right? It is the physical hardware constraint. The dispatch highlights Broadcom's third quarter numbers for financial year 2026. And they are, well, they're eye watering. They really are. $29 .59 billion in total revenue. But the critical line item is their AI semiconductor revenue. $16 .7 billion for the quarter. Wow. Yeah. That is up 221 % year on year. When you see a 221
% increase in a hardware sector, you are looking at the bottleneck of the entire industry. The absolute bottleneck. Because Broadcom makes these specialized networking chips and custom silicon that allow tens of thousands of GPUs to talk to each other in a data center. That $16 .7 billion is an abstract wealth. It represents physical capacity being built out at a frantic pace. Like actual metal and wire. Yes. It represents copper wiring, immense power grid consumption, and massive
liquid cooling infrastructure. And for you listening right now running a business, this isn't just Wall Street trivia. The bill for the AI boom always comes due. Someone has to pay for all that copper and cooling. Exactly. When that much capital is required just to keep the servers running, it trickles down. It means your general cloud hosting costs are going to creep up. It means memory pricing spikes. The specific tier of enterprise software you rely on gets a little more expensive. You are looking at the source of those downstream effects. Precisely. All that
physical silicon running incredibly hot around the clock is basically baking the next generation of intelligence. Which brings us directly from the hardware to the software it powers. Right. And this is where the dispatch shifts from financial metrics to what feels like, honestly, an existential shift in capability. OpenAI just shipped GPT -6, codenamed Astra. The big one. Yeah. And the generational jump in capability here is doing something entirely unprecedented to cybersecurity. Astra is the first model to cross the critical
threshold under OpenAI's own internal preparedness framework. What's fascinating here is how that critical rating is actually quantified. The security metrics they publish represent a profound leap. I mean, they are wild. They are. On a standard industry benchmark called ExploitBench, Astra scored 100%. A perfect score. A perfect score. But the metric that should make every security professional pause is the harder test. They unleashed Astra on fresh vulnerabilities. Specifically, 20 high -severity bugs found in Google Chrome's
JavaScript engine just between June and August of this year. So totally new stuff. Exactly. These are novel, undocumented flaws. And Astra scored 39 % on that test. To put that in perspective, the previous model scored 5 .5%. That is a massive jump. Huge. And during the evaluation, OpenAI says the model actually found and exploited two previously unknown zero -day vulnerabilities entirely on its own. Just unbelievable. We are talking about it reading the underlying logic of a system and identifying a structural flaw
that humans hadn't even categorized yet. It is not just regurgitating code it memorized from Stack Overflow anymore. It is demonstrating active reasoning. It is structural analysis. at machine speed. Now, to be clear, the consumer version of Astra that was actually shipped to the public has heavily restrictive guardrails. Oh, of course. Right. It will refuse to write proof of concept exploit code if you ask it to. But the ship product isn't really the point. The point is the capability exists in the wild now. And it represents a massive shift in the direction of travel. Historically, there's always been a window of time between
a software vulnerability being discovered and a malicious actor figuring out how to weaponize it. Like a grace period. Exactly. Five years ago, that window was weeks. Lately, it has been days. This technology threatens to shrink that weaponization window to absolute zero. If an AI can analyze a patch, understand the vulnerability it fixes, and generate an exploit in seconds, well, human defenders simply cannot respond fast enough. Here's where it gets really interesting, though. Because on the exact same day they announced
this terrifying new capability, OpenAI also announces a $1 billion subsidy for defensive AI tooling. Oh, the timing on that. Right. It is aimed at critical infrastructure, local governments, and open source maintainers. The details are a bit murky, you know. It looks like a six -month run prioritizing U .S. infrastructure. But the timing is just wild. OpenAI is heavily subsidizing cyber defense at the exact same time they're building the ultimate cyber offense tool. It's quite the
paradox. It feels a bit like selling the locks and the lock picks at the same store, doesn't it? It really highlights the... unavoidable dual -use nature of artificial intelligence. OpenAI recognizes that by pushing the frontier of offensive capabilities forward, even if they keep the raw model internal, they are inherently destabilizing the broader security environment. Yeah, they have to know that. They do. The billion dollar subsidy is essentially a public acknowledgement that human defenders can no longer keep up with machine speed attacks unless they are equipped with machine speed defenses. But the irony here
is that we are worrying about Astra weaponizing zero days in some hypothetical future. While right now, today, the infrastructure we built to host these AIs is practically leaving the door wide open. It is a mess out there. This shrinking threat window is happening today, and the first casualties are the very AI systems businesses just rushed to build. Yes, the so -called AI plumbing is leaking bad. Leaking everywhere. The dispatch highlights an urgent alert from CISA. That's the American Cybersecurity and Infrastructure Security Agency. In a single day, CISA added
seven actively exploited vulnerabilities to their catalog. Seven in one day. And we should clarify. The CISA catalog isn't a list of theoretical bugs. If something lands on that list, it means someone is actively using it to break into systems right now in the real world. And four of those seven vulnerabilities were in AI and DevOps plumbing. We are talking about integration tools like Kestra, JFrog, Starlet, and an authentication flaw in Barry's Letellum proxy. A lot of people rely on those. They really do. And if you aren't familiar
with these tools, they are essentially the glue. They move code around. They route your prompts to different language models. They handled the API keys. Microsoft observed attackers exploiting these specific flaws as early as June. And what the attackers were doing with that access is critical to understand. They weren't just, you know, asking the AI to write bad poetry for free. Right. If only it were that harmless. I know. They were deploying reverse shells, which essentially gives the hacker a hidden remote command line directly into the server. They were dropping crypto miners to steal computing power. And most
dangerously, they were specifically targeting the database tables that hold provider API keys. If we look at how most businesses have operated over the last 18 months, there has been this mad rush to stand up AI gateways and proxies to connect company data to these large language models. But they are treating these AI gateways like simple, isolated internal websites. Which is a huge mistake. It is. Treating a highly connected AI gateway like a simple, isolated website is like leaving the master key to a bank vault under
the front doormat because you assume nobody will look there. I mean, that gateway is connected to everything. Your billing API, your proprietary databases, your internal file systems. If we connect this to the bigger picture, Microsoft's explicit recommendation in this dispatch addresses exactly that architectural failure. Businesses must monitor these AI workloads according to their control plane role. The control plane. Yeah, the control plane is the nervous system of your network. Your AI gateway has the keys to your kingdom. you must treat it with the exact
same security rigor, intense logging, and strict access controls that you would apply to a primary domain controller. Which makes total sense. But the dispatch takes a really bizarre turn right after this. Oh, yes, it does. It shifts from these hyper -complex, futuristic AI architecture problems down into the absolute mud of basic web plugins. It's jarring, honestly. It is. Because while we have frontier AI models finding novel zero days, Businesses are still getting completely compromised because they won't update their WordPress
sites. It is the great paradox of modern cybersecurity. The most advanced multi -billion dollar threats share the exact same stage with the most basic human complacency. Let's look at this Elementor Pro vulnerability. If your business runs a WordPress site, you almost certainly know Elementor. It is a massive page builder plugin. Very popular. Super popular. Well, version 4 .2 .2 had a critical arbitrary file upload vulnerability. As of early September, roughly two -thirds of Elementor's
10 million active installations were still running the vulnerable version. Two -thirds. That is millions of exposed sites. Millions. The security firm WordFence had to block over 190 ,000 exploit attempts almost instantly. The mechanism of the exploit itself is a perfect example of how a tiny logical flaw can completely compromise a system. Right, it's a flaw in the validation loop. Explain how that works. Okay, so normally when a user... files to a website, the system checks them one by one to ensure they're safe, checking that it is actually a standard JPEG
image file, for example, and not some malicious executable script. Makes sense. But in this specific version of Elementor, if the validation loop hits an empty upload slot, it throws an error and just stops checking entirely. It skips validation for all the remaining files in the queue. Yes. So what does this all mean? It's like a bouncer at a club checking the ID of the first person in line, finding out the ID is completely blank, getting confused, throwing his hands up, and just letting the next 50 people walk right past him into the club, unchecked. It is a brilliant
analogy. Because what the attacker does is submit an array of files. The first slot is deliberately left empty. The system hits it. gets confused, and stops the security check. And then the bad stuff gets through. Right, because the second slot contains a malicious PHP payload. Yeah. Because the validation stopped, that PHP file, which is essentially a script that the server will execute, gets written directly to the server's disk, completely unchecked. And once that PHP file is on the disk, it is game over. The attacker just navigates to that file's web address, the
server executes it, and they have complete remote control over your entire website. The dispatch has a very clear, urgent action item here. If you use Elementor, you need to go check your web content slash upload slash Elementor slash forms directory. Yes, immediately. If there is a PHP file sitting in there, it is not a false positive. You have a backdoor and your site has been compromised. And this wasn't even the only mundane failure of the week. There was also a critical SQL injection flaw in Sangoma's Switchvox.
Which is a VoIP phone system, right? Voice over IP, essentially the system routing all the office phone calls over the internet. Correct. And this flaw had a CVSS score of 9 .3 out of 10. Wow, 9 .3. Yeah. The Common Vulnerability Scoring System ranks how severe and easy to exploit a bug is. A 9 .3 essentially means anyone with an internet connection and basic tools can likely break in without needing any specialized skills or prior access. So it's basically an open door. Pretty much, Sangoma patched this incredibly
dangerous flaw in mid -July, but attackers were seen exploiting it in the wild at the end of August. That is a 47 -day window, 47 days where the patch was sitting there available, and most small businesses simply did not apply it, leaving their entire internal communication system exposed to the open internet. It happens all the time. And then, of course, to cap it all off, we have Google shipping their sixth zero -day Chrome patch of the entire year, fixing a flaw in the V8 JavaScript engine. This raises an important question. How do you apply a Chrome update? The
browser literally does it automatically in the background. The fix is already downloaded and sitting there on most users' computers. So we have built these trillion -dollar intelligence systems, and the security of our network still comes down to... Did you close the window and turn it back on again? It forces us to confront the human element in all of this. The Chrome update is waiting for a user to simply close a browser window that they probably haven't closed since March because they don't want to lose their tabs. We're all guilty of tab hoarding. Oh, absolutely. But we can debate. billion dollar AI defense
subsidies and autonomous threat hunting frameworks all day long. The single highest value, lowest effort security habit available to any office right now is just telling your staff to close their web browser completely once a week. It's so simple. It is. The attack surface of any business is ultimately managed by humans. Humans who are tired, distracted, or simply unaware that leaving a browser open for six months is a critical security risk. Let's quickly recap the journey we have been on today because it really is a wild ride.
We zoomed way out to see the billions of dollars pouring into the physical foundation of AI through companies like Hugging Face and Broadcom's massive hardware build -out. The macro view. Right. Then we watched as AI crossed a critical, terrifying threshold with GPT. The overarching takeaway
for anyone listening is that knowledge is only power if you actually apply it. Understanding the macro trends like the shrinking weaponization window is vital for your long -term strategy. But the micro actions like patching the VoIP system are what keep your business alive today. So for everyone listening, your homework is crystal clear. Go check your WordPress plugins right now and write down your AI model dependencies before the end of the day. But before we wrap up, I want to leave you with one final thought to mull over, building on everything we have explored in this dispatch. If we look at the
trajectory we have mapped out today, as AI capabilities shrink that vulnerability window to near zero, And as models like GPT -6 begin to find and exploit novel zero days entirely on their own, are we rapidly approaching a threshold where humans are completely removed from the cybersecurity loop? That is a terrifying thought. We might soon be facing a reality where AI perpetually attacks and patches AI infrastructure at speeds that human cognition cannot even perceive. We could be looking at the dawn of cyber warfare
as a machine -only domain, where we are just the bystanders, entirely reliant on the structural integrity of the code. that is a staggering thought a silent invisible war happening millions of times a second in the background just to keep our browser tabs from crashing thank you all for joining us on this deep dive into the architecture of our very weird very fast moving future good luck out there in the digital wild and seriously go restart your browser we'll catch you next time
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Japanese home goods company Iris Ohyama (online store operator Iris Plaza) found a suspicious file in an external...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Australian online retailer Who Gives A Crap told customers on 7 November 2026 that their information was accessed...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Last week in tech, 7 to 14 November 2026. Microsoft's November Patch Tuesday (10 November) fixes well over a...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Chainalysis disclosed on 30 October 2026 that an external actor using "offensive agentic techniques" accessed...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! NSW Health reported unauthorised access to its Oral Health Fee For Service Scheme website on 29 October 2026;...