Last Week in Tech – Anthropic vs Alibaba, OpenAI's Own Chip, and an Emergency Cisco Patch
19 minAudio
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Your Monday catch-up, with a security lens. This week: Anthropic takes an extraordinary allegation to US lawmakers, accusing Alibaba's Qwen team of the largest-ever "distillation attack" on Claude. OpenAI and Broadcom unveil "Jalapeño," OpenAI's first custom AI chip. Qualcomm spends nearly $4 billion buying Modular to chip away at Nvidia's CUDA lock-in. And CISA orders an emergency patch of an actively-exploited flaw in Cisco's phone-system software. Wondering whether that Cisco flaw affects you? Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: CNBC; OpenAI; Bloomberg; CISA; BleepingComputer; The Hacker News; Help Net Security.
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.
Imagine finding out that the biggest threat to a multi -billion dollar AI isn't some highly sophisticated nation state super virus. Right. But instead it's just a quiet army of like 25 ,000 bots asking it to write cunt for pennies. Yeah, it's wild. Or that the key to breaking a trillion dollar hardware monopoly is essentially just a software universal translator. Welcome to this custom deep dive. Today our mission for you listening is to really look underneath the
hood of the AI applications you use every single day and well... examine this massive structural shift that's happening right now. Yeah. We're working from a really fascinating, incredibly dense dispatch today from a technology update. Right. Tech Talks with Kinsoft. Exactly. From June 29th, 2026. Yeah. And I mean, this source covers just one week in the tech industry. Yeah. But it acts like this perfect time capsule. So, okay, let's unpack this. Because reading through this dispatch, you realize really quickly that the, you know, the experimental novelty phase
of AI, it's completely over. Oh, completely. It's done. We are looking at high stakes battlegrounds over intellectual property, massive hardware monopolies, and basically the dawn of automated cyber warfare. It is a really dense week of news. And I think the through line connecting all these seemingly separate events is really just one word, which is infrastructure. Infrastructure. Right. Yeah. Everything we're going to examine today really demonstrates how the very foundation of artificial intelligence is just being violently
shaken up and rebuilt. Yeah, rebuilt from the ground up. Well, we are seeing a complete reconstruction happening on every single level. You know, how the underlying training data is safeguarded, how the physical silicon itself is manufactured, and how the entire security architecture of our networks is fundamentally changing. Well, I want to start with the intelligence itself. Because the Kinshoff dispatch kicks off with this wild story about a new kind of IP theft called distillation. Yes. Distillation. But before we get into the mechanics of this, and I want to make something
explicitly clear to you listening, we are impartially reporting the contents of this source document today. Right. That's a very important distinction to make. Yeah, because the dispatch details accusations made by one company against another. You know, we are not endorsing these viewpoints and we aren't taking sides here. We are simply conveying the claims and the underlying technology that's described in the source material. Absolutely. And it's an important distinction because, well, the claims are explosive. On June 24th, 2026,
the dispatch notes that Anthropic actually went to a U .S. Senate committee. Which is a big deal on its own. Right. And they presented a document accusing operators tied to Alibaba's Quen Lab of executing the largest known distillation attack on their clawed models. The numbers they presented to lawmakers are just, they're staggering. I mean, they alleged roughly 25 ,000 fraudulent accounts were created. 25 ,000. Just to bypass API rate limits. And through those accounts, these operators supposedly ran nearly 29 million
conversations. Over the course of about six weeks, I think. Yeah, six weeks. And the goal here was to cheaply copy, or as they call it, distill Claude's capabilities, specifically targeting its coding, reasoning, and cybersecurity skills to train a rival model. Right. Now, Alibaba completely denies this, obviously, and as the source notes, it hasn't been independently verified. Putting the geopolitical drama aside for a second, let's talk about the mechanics here. Let's do it. Because
I understand using bots to scrape data, right? That makes sense. But wait, if you're just copying the outputs of a model, how does a smaller competing model actually learn reasoning? That's a great question. Isn't it just like parroting answers without understanding the underlying logic? Well, what's fascinating here is the underlying mechanism of how synthetic data transfer works in modern neural networks. You aren't just scraping the final answers. When an operator runs a distillation attack on this massive scale, they're actually prompting the target model to generate what we
call chain of thought output. Chain of thought, right. Right. They are asking the model to show its work step by step before arriving at the solution. Ah, I see. So instead of just saying, you know, write this final line of code. They're asking it to explain why it chose a specific variable or a specific algorithm in the first place. Precisely. To use an analogy, think of the frontier model, in this case, Claude, as a master teacher. Okay. A master teacher. Yeah. A teacher who has already spent years reading every single textbook in the world. A distillation
attack is essentially like forcing that master teacher to sit down and write 29 million highly detailed step -by -step study guides. Wow. Okay, so then the smaller student model. The student model doesn't need to read the entire library of human knowledge anymore. Right. It just reads the study guides. Exactly. It trains directly on those pristine, highly structured study guides. It learns the logical pathways, the reasoning itself, without the massive computational overhead of having to figure it all out from scratch. That is just wild. I mean, it completely bypasses
the billions of dollars and, what, months of continuous data center compute required for original training? Oh, absolutely. You are literally weaponizing your competitor's multi -billion dollar machine to train your own machine. Which is exactly why this is considered asymmetrical warfare. Yeah. The cost to defend against this is incredibly high. Because think about it, to the frontier model servers, this attack just looks like normal, albeit heavy, user traffic. Right, because it's distributed across thousands of Sybil accounts.
Just looks like a bunch of users chatting. Exactly. And taking this to a U .S. Senate committee signals something profound here. It shows that the traditional tools we have for protecting intellectual property. Like copyrights and patents. Yeah, copyrights, patents or even basic terms of service. They are completely failing in the AI age. So, OK, if software IP is suddenly impossible to defend from distillation like your digital moat is basically evaporating. Is that why these same companies are suddenly pouring billions into proprietary
silicon? That's exactly it. Are they basically trying to build a physical moat because the digital one just totally failed? That is the exact strategic pivot the entire industry is making right now. Wow. If you can't easily lock down the software weights of your model, well, the next logical step to maintain your edge and your profit margins is to aggressively control the physical hardware it runs on. And that pivot leads us directly to the second major theme in the Kinsoft dispatch, which is this. Yes. Two massive stories landed on the exact same day and both squarely targeted
NVIDIA's absolute dominance over the AI silicon market. It was a big day. Yeah. First up, OpenAI and Broadcom unveiled Jalapeno, which is OpenAI's very first custom in -house AI inference processor. But here's where it gets really interesting. The dispatch reports that this chip went from its initial design phase to tape out. In just nine months. Right. Nine months. And they're planning to deploy it in gigawatt scale data centers by the end of the year. I have to push back on this timeline a bit. Sure. Nine months
from design to tape out. That feels like it breaks the laws of physics and semiconductor manufacturing. Yeah. Even, you know, insanely well -funded teams take years to get silicon ready for the foundry. Yeah. Usually they do. How is that even mechanically possible? Well, it sounds impossible if you assume they started from a blank whiteboard. OK. They didn't. To pull off a nine month tape out. OpenAI heavily leveraged Broadcom's existing pre -verified IP blocks. Ah, I see. Think of it like building a house with prefabricated walls and plumbing
rather than mixing the concrete yourself on site. Broadcom already has industry -leading SUEs, which are these serializer, deserializer blocks that handle high -speed data transfer between ships. Right, the networking, basically. Exactly, and they already have proven memory controllers. So OpenAI essentially brought the specific matrix math architecture they needed for their AI, and Broadcom... just wrapped it in their existing foundational tech. So they basically snapped these high -end Lego pieces together. Pretty
much, yeah. But even then, I mean, why go through the massive headache of building this when they already have access to NVIDIA's top -tier hardware? Like, they already have the best stuff. It really comes down to the architecture of inference versus training. Okay, break that down for me. So... NVIDIA's flagship GPUs are incredible general -purpose AI accelerators. They are built to handle the intense computational precision you need for backpropagation during the training phase. Right, when the model is actually learning. Yes.
But Jalapeno is an inference -only chip, meaning when a model is just answering user prompts in production, you don't actually need that massive compute density. What you need at that stage is memory bandwidth. You need to move weights in and out of SRAM as fast as possible with the lowest possible latency. Okay, so it's a completely different job. Exactly. By designing a chip strictly for inference, OpenAI strips away all that silicon real estate dedicated to training and replaces it with pure memory throughput. Which means they
run their models vastly cheaper. faster and cooler than using general purpose hardware. Absolutely. Okay. It makes total sense for them to carve out independence on the physical silicon side. But then literally on that exact same day, we get the second half of this rebellion. The software side. Right. Qualcomm confirmed it's acquiring an AI software startup called Modular for over $3 .9 billion. A huge acquisition. Massive. And the dispatch says Modular created a hardware agnostic runtime that attacks NVIDIA's software
lock -in. Which we know is CDA. Yes, CDA has been the gold standard. But again, let's get into the mechanics here. People always throw around the phrase universal translator for software like this. But how does a hardware agnostic runtime actually decouple the code from the chip? Well, if we connect this to the bigger picture, we really have to look at the compiler level. For over a decade now, if a developer wrote AI code in a framework like, say, PyTorch, the absolute easiest way to make it run fast was to use NVIDIA's proprietary CUDA libraries because they are highly
optimized for NVIDIA's specific GPU architecture. Right. It created a massive walled garden. You basically had to use their hardware if you use their software. Exactly. It's a complete ecosystem lock -in. Modular tackles this by utilizing an intermediate representation layer, specifically something called MLIR, or multi -level intermediate representation. So it basically sits between the high -level Python code and the actual metal of the chip. Exactly that. Instead of compiling directly to CUDA, Modular's engine takes the
AI workload, breaks it down into this intermediate mathematical representation, and then uses custom compilers to translate it perfectly into the native instructions. set of an AMD chip, an Intel accelerator, or even, you know, a custom chip like Jalapeno. Wow. It mathematically decouples the software layer entirely from the physical hardware layer. So what does this all actually mean for the listener? I mean, we have this incredible pincer movement happening here. Yeah, a two -pronged attack. Jalapeno attacking NVIDIA's hardware dominance by proving companies can build custom
silicon rapidly. And then Modular attacking the software dominance by breaking that CDA lock -in. Mm -hmm. Why should a business leader or just an everyday user care about this highly technical silicon war? Because this fierce competition is the exact mechanism that will drive down the operational costs of the entire digital economy. It always comes back to cost. Always. When developers aren't forced to buy one specific brand of hardware, and when companies like OpenAI aren't paying massive premiums for general purpose inference
compute, the unit economics of AI just... plummet. Right. For you, the listener, it means AI integrations in your enterprise software, your customer service tools, your daily workflows. They will all become significantly cheaper and exponentially more ubiquitous. You are literally watching the monopolistic tax being stripped out of the ecosystem in real time. OK, so we are rapidly building out gigawatt scale data centers for these custom chips. Yes. And we're deploying massive new software compilers to route these workloads. Oh. But as we pour
the concrete for these data centers and write millions of lines of new orchestration code, the attack surface for bad actors expands massively. Oh, without a doubt. And that brings us to the final and frankly, the most urgent section of the Kinsoft Dispatch today, which is the cybersecurity warnings. It is just the unavoidable physical reality of infrastructure expansion. You cannot scale data centers to the gigawatt level and rewrite your entire software stack without introducing new vulnerabilities. And the vulnerability highlighted in this dispatch is just profoundly ironic. It
really is. I mean, we just spent 20 minutes talking about autonomous reasoning models and hyper advanced AI silicon. But on June 25th, the American Cyber Agency, CISA, issued an urgent patching directive for a phone system. Yep. iPhone system. Specifically, Cisco Unified Communications Manager. The bug is cataloged as CVE -2026 -2030. It's an unauthenticated flaw that let attackers write files directly to the underlying system, and it was being actively exploited in the wild. Actively exploited, which
is why the June 28th deadline was so hard. Right. But I have to admit some confusion here. Why do sophisticated threat actors, who could theoretically be targeting these massive AI clusters, care about a boring legacy office phone system? Well, this raises an important question about network topology and something called lateral movement. Attackers don't usually try to breach the most heavily guarded fortress door, right? They look for the forgotten side window. The easy way in. Exactly. A unified communications system is essentially a server that sits right on the edge of your
network. It has to connect to the public internet to route voice traffic. Sure. But it also has deep, trusted connections to your internal enterprise network to access things like user directories and databases. Ah. So it's basically a bridge between... the wild west of the Internet and the safe internal corporate network. Yes, exactly. And because it's an appliance that just quietly works in the background, IT departments often treat it as a set it and forget it device. Right. It just gets pushed to the bottom of the patching queue. Right. And when an attacker exploits an
unauthenticated file write flaw on that specific appliance, they aren't just messing with your voicemails. They are dropping a web shell. Oh, wow. They gain root access to that server. And from there, they pivot. They use that trusted appliance to move. laterally inside the network, targeting the domain controllers, deploying ransomware or, you know, exfiltrating data. Yikes. So the grounded practical takeaway for you listening is this. You absolutely must treat every single piece of internet -facing enterprise software, especially the mundane legacy appliances like
a phone system, as a critical infrastructure vulnerability. Which honestly feels a bit overwhelming. It is. I mean, the sheer volume of code in modern enterprises makes manual patching feel like trying to empty the ocean with a teaspoon. Exactly. But the dispatch does introduce a massive shift in defensive capabilities to counter this, right? It does, yeah. Because that same week. OpenAI expanded its Daybreak cybersecurity program. They released a specialized security -focused AI model to vetted defenders. And the pitch here
isn't just an AI that flags anomalies, right? No, it goes much further than that. It includes tools for automatically finding and patching vulnerabilities in code at scale. So instead of a human security analyst endlessly hunting for zero days, the AI does static analysis, understands the semantic context of the code, and basically writes the patch itself. And this brings us to the terrifying duality of autonomous agents and cybersecurity. Right. Because the concept of automated patching, where a system discovers a buffer overflow and writes the mitigation code
in literally milliseconds, it's structurally necessary for defenders to keep up at this point. Yeah, because humans are just too slow. Right. The time to patch window is basically shrinking to zero. Yeah. But the dispatch cautions heavily about the dual use nature of this technology. Right. Because an AI model that is sophisticated enough to read millions of lines of code, find a deeply hidden logic flaw and understand how to patch it. Yes. Is also, by definition, sophisticated enough to exploit it. Exactly. We are transitioning
right now into a machine speed arms race. The same static analysis capabilities that empower the Daybreak program to secure an enterprise can be entirely inverted by threat actors. Wow. They can use it to discover zero -day vulnerabilities in proprietary software before the vendor even knows they exist. That is just wild to think about. We are moving away from the era of human operators manually probing networks into this era where automated AI tools are simultaneously defending and attacking infrastructure at speeds
human analysts simply cannot process. Wow. OK, let's take a breath and just bring all of this together for you. We started this deep dive looking at a single dispatch from June 2026. But it's not just a laundry of tech news. It is a high definition snapshot. Absolutely. Right. Yep.
And finally, we looked at the massive security fallout of this expansion, balancing the urgent reality of lateral network attacks via Cisco phone systems against the futuristic machine speed promise of OpenAI's Daybreak program. It's a lot to take in. And as we consider the architecture of this new era, I want to leave you with a final thought to mull over, building on that last point about security. Okay. We discussed the duality of AI and cybersecurity and that shrinking time -to -patch window. Well, if AI systems can automatically find and patch vulnerabilities at scale, as Daybreak
promises, what happens when two opposing AIs, one designed to relentlessly attack and one designed to autonomously defend and engage in an automated cyber war at speeds no more than a second, No human can monitor. Oh, wow. Will our enterprise networks and the basic Internet infrastructure we rely on every day simply become passive battlefields for autonomous agents fighting invisible millisecond wars? That is a wild and honestly somewhat chilling thought to leave on. The landscape is entirely new and the fundamental infrastructure of the digital world is being rewritten as we speak.
We hope this deep dive gave you the mechanistic context you need to truly understand these shifts, because taking the time to explore the how and why is exactly how you stay ahead of the curve. Thanks for joining us today. Stay curious and stay patched.
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;...