Last Week in Tech – Anthropic's $900B Round, AI Guardrails Stripped in Minutes, and Trellix's Source-Code Breach
23 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 is reportedly closing a $30B round at a $900B-plus valuation and projecting its first-ever quarterly operating profit - even as SpaceX filings reveal eye-watering compute bills. The politics heat up as President Trump abruptly cancels an AI executive order. A worrying safety finding shows researchers stripping the guardrails off major AI models in minutes. And on the breach desk: security vendor Trellix has attackers in its own source code, drug-delivery manufacturer West Pharmaceutical recovers from a crippling attack, and education-sector attacks jump again. Concerned your security tools themselves could be a target? Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: Tech Startups; BuildFastWithAI; Bloomberg; Comparitech; cm-alliance.
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.
Welcome to the Deep Dive. Today, we are exploring what I can really only call this, you know, wild, dizzying paradox in the current tech landscape. Yeah, it's a lot to wrap your head around, honestly. It really is. So we have a single, just densely packed source for you today. It's the June 1st, 2026 edition of Kinsoft Tech Talks. And the report is called The AI Economy and Security Frontiers. Right, which sounds very formal, but the contents are just... They're wild. Completely. And our
mission for this deep dive is to basically navigate this incredibly stark contrast. Like we are going to take you from the staggering just money printing heights of this current AI boom right down to the, frankly, alarming vulnerabilities in the cybersecurity foundation that's supposedly holding it all up. And that is really the crucial dynamic to understand right now. My goal today is to help you, the listener, connect the dots between that. you know, unprecedented economic acceleration and the actual security fragility underneath it all. Because it feels like two completely
different realities, right? Exactly. Because when you read the financial headlines, you see one very specific version of the future. But then, well, when you look at the infrastructure and the daily security realities, a completely different, much more precarious picture emerges. I have to say, just looking at the sheer scale of the information in this Kinsoft report. It is a lot to process. I mean, the numbers, the speed of development, the geopolitical stakes. It really requires a total reframing of how we think about technology. It does. It absolutely
does. And that is exactly why we are breaking it down today, because understanding this landscape, it isn't just for like. tech insiders anymore. You know, it dictates the environment you are going to be operating in, regardless of what industry you're actually in. Right. OK, let's unpack this, because the momentum of the tech world right now is, I mean, fundamentally staggering. Yeah, it's moving at a pace we haven't really seen before. So the report kicks off with Anthropic, the company behind the Claude AI assistant. And according to the source, they are closing this 30 billion dollar funding round. 30 billion.
Which is just a massive injection of capital. And that puts their valuation north of $900 billion. $900 billion. Yeah, a valuation that basically places them among the most valuable entities in human history. And they achieved it in like a fraction of the time it took legacy tech giants to reach similar heights. And it's not just... you know, purely speculative value based on some future promises. The Kinsoff report notes that Anthropic is actually on track for its first ever quarterly operating profit. Right. They
are projecting nearly $11 billion in revenue for the second quarter alone. $11 billion in a quarter. And at the exact same time, OpenAI is reportedly preparing a confidential filing to go public. I mean, the amount of capital just swirling around this space right now is immense. It really is. But then, you know, you look at the expense side of the ledger. Oh, man. Yeah. Which is arguably the most telling part of this entirely economic equation, really. Because in SpaceX's recent filings, it was revealed that Anthropic is paying SpaceX over a billion dollars a month for GPU compute. A billion dollars. Every
single month. It's just it's hard to fathom. I want to offer an analogy here because I keep thinking about like the Industrial Revolution. That's a good comparison. Yeah. Right. You look back at that era and you had these massive tycoons building the giant, highly profitable factories that just fundamentally changed society. The Carnegie's, the Ford's. Exactly. And right now, Anthropic and OpenAI, they are building the modern equivalent of those massive factories. But SpaceX, SpaceX. is the one selling them the coal to keep
the furnaces burning. Selling them the raw power, essentially. Yes. And they're selling them a billion dollars worth of coal every single month. But my question to you is like, how on earth is that sustainable? I mean, is a $900 billion valuation just a bubble when your overhead is a billion dollars a month? Well, what's fascinating here is what that dynamic actually reveals about the mechanics of the AI economy itself. It is entirely unprecedented. When you ask if it's sustainable, you really have to understand why those costs are so high in the first place. Right,
because it's not like normal software. No, this isn't like traditional software at all. How so? Because, I mean, if I build a traditional sauce application, my cloud hosting bill doesn't look anything like that. even if I have millions of users? Because traditional software is essentially deterministic. You write the code, you compile it, host it on a server somewhere, and when a user clicks a button, the server just executes a pre -written set of instructions. So the computational cost to run it for one user versus a million
users, it scales incredibly efficiently. Right. It's just retrieving things. Exactly. Generative AI fundamentally breaks that scaling model. So every time someone queries an AI, it's like a much heavier lift for the system. A massively heavier lift. It's a process called inference. So every single time you ask a large language model a question, it isn't just retrieving a static file. It's actually thinking. Well, calculating. Right. Those server farms have to run these incredibly complex probabilistic calculations across hundreds
of billions of parameters in real time just to predict the next word and then the next word and the next. And that requires an immense amount of continuous electricity and highly specialized processors, you know, the GPUs. You are essentially renting a supercomputer for every single individual query. Which totally explains the billion dollar SpaceX bill. Yeah, the compute power required just to keep the lights on and answer user prompts is astronomical. And that's not even talking about the compute required to train the next generation of models. Right. That's a whole other
massive expense. Exactly. So to answer your question about sustainability, a $900 billion valuation means the market truly believe these models will restructure the global economy so thoroughly that a billion dollar monthly overhead. will eventually just look like a bargain. Like a drop in the bucket. Right. But in the short term, those staggering bills force these companies to operate at a breakneck, just relentless speed. They can't afford to slow down. They really can't. They have to keep growing, keep deploying, and keep monetizing at maximum velocity just to outrun
their own infrastructure costs. Which naturally leads us to the next major point in the source material. And here's where it gets really interesting, because... If these companies are building the future at this blistering speed due to financial pressure, you have to ask, you know, who is directing traffic? Who's in charge, yeah. Right. Who is setting the rules of the road? And according to the Kinsoff report, the answer seems to be that policy is entirely unsettled. The policy vacuum is, I'd argue, a defining characteristic of this current era. Yeah. The report highlights
this very specific event. President Trump abruptly canceled the signing of a major AI executive order. Right. Just pull the plug on it. And the explicit reason provided for canceling it was that it risked undermining America's competitive edge. And just to be clear, looking at this strictly as a classic technological dilemma, completely setting aside politics, this feels like the ultimate double -edged sword. Oh, absolutely. Because on one hand, you obviously want to win the global economic race. But on the other hand, by intentionally removing the speed limits, aren't we kind of
driving blind into this incredibly powerful technology? Well, if we connect this to the bigger picture, what you are observing is essentially a classic game theory scenario. But it's playing out on a geopolitical scale. Right. It is that fundamental historical tension between regulation and innovation. But now it's amplified by the sheer speed of AI development. Nations are acutely aware that AI is not just another software product. It's way bigger than that. It is viewed as the foundational
technology of the next century. much like the electrical grid or the Internet itself. I understand the desire for dominance, obviously, but surely establishing some basic federal -level guardrails would stabilize the market. Like, doesn't a total lack of rules invite disaster for the companies who are actually trying to build in this space? I mean, from a purely domestic standpoint, yes. Clear regulations provide market certainty, which businesses love. Right. But the issue here is international parity. Okay. If you heavily regulate your domestic AI industry to ensure it is perfectly
safe, thoroughly tested, ethically constrained, you inherently slow down the development cycle. Because all that testing takes a lot of time. Exactly. And in a vacuum, that's the responsible thing to do. But we don't live in a vacuum. The calculation being made is that if you slow down your own tech giants, international competitors who, frankly, do not share your regulatory concerns will just race ahead. And they will capture all the economic benefits, the intellectual property, and, I mean, the military applications. Precisely.
The state apparatus is essentially looking at the landscape and deciding that the risk of falling behind is fundamentally greater than the risk of moving too fast. That is exactly why the executive order was canceled, to maintain that competitive edge. So the immediate takeaway for you listening to this deep dive is that you cannot wait for the government to figure out the rules of AI. You're on your own. Pretty much. You have to navigate this landscape under the assumption that federal policy will remain entirely unsettled for the foreseeable future. Which means the responsibility
for safety and security, it doesn't lie with some regulatory body. It falls entirely on the companies themselves. Like, we're relying on the AI models to just police themselves. And that transition of responsibility, that leads us directly into the most concerning technical revelations in the Kinsoft report. Yeah, because if the government is hitting pause on regulations, we are placing this immense amount of trust in the internal defenses of these commercial tools. But the report reveals that those internal defenses are, I mean, shockingly fragile. They really
are. It says researchers reported that they could remove the safety guardrails from major AI models in a matter of minutes. Just minutes. Yeah, it's wide. And once those guardrails were removed, the models just happily answered prompts they were explicitly supposed to refuse. It completely shatters this illusion of inherent model safety that is so often marketed to the public. Honestly, I really have to push back on the assurances we constantly hear from the tech industry. They tell enterprise customers that these models are aligned, that they have these strict safeguards,
that they won't generate malicious code or expose sensitive corporate data. Right. That's the sales pitch. But if it only takes a team of researchers a few minutes to break those rules, are these guardrails essentially just security theater? To me, it sounds like putting a flimsy piece of yellow caution tape in front of a massive bank vault. You can issue a press release saying the vault is protected, but like anyone can just duck under the tape. This raises an important question, though, which is what exactly is a guardrail in the context of a large language
model? Good question. I think the general public and honestly, even many IT professionals, they imagine a physical barrier or like a hard coded firewall. Right. Like an if then statement. Exactly. They imagine a line of code that literally says if user asks for X, block the request. Yeah. But that is simply not how. neural networks operate. How do they operate then? Because clearly the word guardrail is doing a lot of heavy lifting here. Yeah, it really is. A guardrail in an LLM is essentially just another layer of training data. Wait, really? Just data? Yeah. It is a
set of statistical weights that biases the model toward generating a refusal like saying, you know, I cannot help you with that when it recognizes certain patterns of words. Oh, I see. It is soft instruction layered over a probabilistic engine. And the fundamental problem with software instruction that operates entirely on human language is that human language is infinitely flexible. Right, of course, because there are a million different ways to ask the exact same question. Exactly. So clever researchers or malicious actors, they
use what are called adversarial prompts or jailbreaks. Right, we've heard of jailbreaks. Yeah, and they don't hack the code, right? They logically trap the model. They use complex scenarios or role -playing prompts or weird hypotheticals to confuse the model's pattern recognition. So they trick it. Basically. Because the model is just predicting the next most likely token based on its training. If you can structure a prompt that makes a dangerous response statistically more likely than a refusal response, the guardrail just dissolves. You essentially just talk the math out of being safe. You bypass
the statistical bias exactly, which is why there is a massive foundational difference between a model's built -in safety rules and actual corporate data security. Okay, how so? Well, the model's internal safety is designed to stop it from generating generally harmful content, like bad instructions or toxic speech. It is not a bespoke, rigid security architecture designed to protect your company's proprietary data from a targeted extraction attempt. Wow. Which brings us to the very... explicit advice Kinsoft gives to businesses in this report.
They state, if you let staff use AI tools, assume guardrails can be bypassed. Put your own policies and monitoring around how tools are used with company data. And that is the crucial pivot every single organization needs to make in their threat modeling right now. You have to assume it's vulnerable. Yes. You cannot trust the AI's internal safety features to protect your environment. If you are integrating these tools into your daily workflows, you have to operate under the baseline assumption that the AI will at some point behave exactly as a malicious actor or an accidental insider
wants it to. That's terrifying, honestly. It is, which is why you must rely on your own internal monitoring, your own data loss prevention tools and strict traditional access controls. OK, so just to summarize the logic here, because the AI's internal defenses are easily bypassed through language manipulation, we have to fall back on our traditional cybersecurity defenses. We have to basically trust the enterprise security tools we've relied on for the past decade to catch any malicious behavior. Yes, that is the standard.
But as the report details in its final section, that foundational layer of legacy cybersecurity, it's cracking under pressure as well. And this is the part of the report that completely changes the threat model for every IT department. It's massive. Because the vulnerability isn't just limited to brand new experimental AI tools where we kind of expect some rough edges. We are dealing with the reality that our established cybersecurity infrastructure is being deeply compromised. The report highlights that a major trusted security
vendor, Trellix, was breached, but they didn't just lose customer data, which is bad enough. Attackers gained access to the source code behind Trellix's own security products. And in the cybersecurity world, unauthorized access to a vendor's source code is a worst case scenario. It is a fundamental breach of trust. So what does this all mean? Let me offer an analogy to make this kind of abstract threat concrete for anyone listening. Go for it. Imagine you own a highly secure facility, right? And you hire a premier security company
to install these unpickable, state -of -the -art locks on all your doors. You feel perfectly safe. Naturally. But then you find out that the company manufacturing those physical locks just had their master blueprint stolen by a highly organized syndicate of burglars. Oh, man. The burglars aren't breaking into your building right now, but they're sitting in a warehouse somewhere. studying the exact schematics of your lock, figuring out precisely where the gears catch and where the microscopic weaknesses are. That analogy captures the severity perfectly. Because when attackers acquire the source code of a security
tool, they aren't just looking for a quick financial gain like a standard ransomware attack. They're playing the long game. Exactly. They are stealing the rulebook. They can compile that proprietary code in their own isolated environments, reverse engineer the detection algorithms, and methodically hunt for zero -day vulnerabilities. Which are the flaws nobody knows about yet. Right. Weaknesses that the vendor doesn't even know exist yet. But doesn't that render the entire concept of legacy cybersecurity obsolete? Like, if the attackers have the blueprints to the defense, why even
bother paying for the software? It doesn't render it obsolete. Because you still need defenses against the millions of automated known attacks that happen every single day. Okay, fair point. But it means legacy cybersecurity is no longer a silver bullet against advanced threats. They can analyze exactly how the security software looks for malware and then design their own malicious code to look like something else entirely. So they know exactly how to disguise themselves. Yes. It is weaponizing. the defense mechanisms against the defenders. It's the terrifying concept
of supply chain and vendor vulnerability. Because in today's digital economy, you aren't just responsible for your own perimeter anymore. No, absolutely not. You could have flawless internal policies. Your employees could be perfectly trained to, you know, never click a phishing link. But if the software you purchase to protect your endpoints is fundamentally flawed because its blueprints were stolen, your entire network is exposed. And the Kinsoft report doesn't just deal in hypotheticals here either. They provide these stark examples of the collateral damage that occurs when these
interconnected systems fail. Real world consequences. Exactly. We aren't just talking about software companies losing data. They highlighted West Pharmaceutical Services, a major manufacturer of drug delivery components. They spent two full weeks recovering from a cyber attack that completely crippled its systems. Two weeks. Two weeks of downtime for a pharmaceutical manufacturer. ripples out. It impacts the global supply chains. It delays shipments to hospitals. It ultimately impacts patient care. It perfectly demonstrates how digital vulnerabilities can just halt physical
infrastructure. And the report also notes that cyber attacks on the education sector jumped more than 50 percent month on month. So we are talking about schools, universities and research institutions being actively targeted now. It just proves that absolutely no industry is insulated from this. No industry is insulated because the underlying digital foundation is a shared resource. Whether you are a critical pharmaceutical giant, a local public school district, or a $900 billion AI startup, you are all relying on a very small,
consolidated handful of security vendors, cloud providers, and software supply chains. We're all in the same boat. Exactly. When one of those foundational pillars is compromised, like a vendor source code being breached, the shock waves hit everyone simultaneously. Which brings us to the vital question every single listener needs to take back to their team tomorrow morning. What happens if one of our trusted security vendors gets breached? It's a complete paradigm shift, really, because we spend all our time running these tabletop exercises asking, what happens if we get breached? Right. You focus on your
own walls. But you have to war game this scenario where the tools you implicitly trust are turned against you or just fail silently. Wow. Do you have redundant detection systems? Can you isolate sensitive network segments if your primary endpoint protection is compromised by a supply chain attack? If the answer is no, you have a massive critical blind spot in your architecture. Which makes Kinsoff's final piece of advice in the report incredibly poignant. They simply state, stay patched, stay skeptical. It is the only rational
operational posture to take in a landscape this volatile. You must relentlessly patch your systems because you have to close the known vulnerabilities. But you must remain deeply skeptical of the software you deploy because you know the unknown vulnerabilities are currently being mapped by highly capable adversaries. It has been a truly wild journey on this deep dive. Let's just briefly recap the sheer scale of the terrain we've covered today. It's a lot of ground. It really is. We started with the absolute economic euphoria of anthropics. Nine hundred billion dollar valuation balanced
against the staggering reality of SpaceX's billion dollar a month compute bills just to keep the inference engines running. Right. The coal for the factories. Exactly. Then we looked at a geopolitical landscape where policy is intentionally left unsettled to win a global race, effectively removing the speed limits on development. We discovered that the internal state. statistical guardrails of these AI models can be bypassed by researchers using adversarial language in mere minutes. Just talking the math out of being safe. Right. And finally, we saw that our traditional defenders,
security vendors like Trellix, are having their own source code breached, leading to devastating real -world consequences for critical industries like pharmaceuticals and education. It is a profound and... honestly, highly volatile convergence of technological trends. And this is exactly why you should care. Whether you are prepping for a high -stakes strategy meeting, trying to protect your own small business network, or you're just incredibly curious about the mechanics of the future, you have to operate with this ghoul
reality in mind. You really do. You cannot just blindly trust the glossy, optimistic brochures of the AI revolution, and you cannot assume your legacy security tool. are impenetrable fortresses, you have to stay patched and you have to stay deeply skeptical of the infrastructure you use every single day. I want to leave you with a final thought to mull over, something that connects the two very distinct halves of our conversation today. Okay, let's hear it. We established that clever researchers can bypass AI safety guardrails using language manipulation in a matter of minutes,
right? Right. We also established that hackers now have the source code to hunt for unseen weaknesses in major security vendors like Trellix. So what happens when those two realities merge? Oh, wow. What happens the day a brilliant guardrail -free AI model is deliberately prompted by a malicious actor to autonomously hunt through that stolen security source code, when an AI that can process millions of lines of code a minute is tasked with finding zero -day vulnerabilities that human
hackers would take years to spot. Oh, man. It's like handing a supercomputer to the burglar who already stole the blueprints to your locks. The pace of discovery is going to accelerate faster than we can even comprehend. We are building the most incredible technological engine in human history, but we desperately need to start paying a lot more attention to the fragile foundation we are laying down underneath it. Thank you so much for joining us on this deep dive. We'll catch you on the next one.
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;...