Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Tech Talks With Kinsoft

Tech Talks With Kinsoft artwork

Last Week in Tech – Big Tech's "We'll Build It For You" Land Grab, a Cheaper Claude, and a CitrixBleed Sequel

Jul 5, 202624 minAudio

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Your Monday catch-up on the week that was in tech, with a security lens. This week: the AI giants pivot from selling software to selling people - Microsoft stands up a 2.5-billion-dollar "Frontier Company" to embed its own engineers inside your business, just two days after Amazon committed a billion to the same idea. Anthropic ships Claude Sonnet 5, a cheaper model built to run autonomous agents. Money keeps pouring into the AI plumbing, with Together AI raising 800 million and Meta reportedly building a cloud business of its own. And on the security desk: a CitrixBleed-style flaw in NetScaler is exploited within a day of its patch, researchers document the first ransomware run end-to-end by an AI agent, and a 19-year-old alleged member of Scattered Spider is extradited to the US. Trying to work out what all this means for your own tech stack and your patch queue? Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: CNBC; TechCrunch; GeekWire; Anthropic; Bloomberg; Business Wire; SecurityWeek; The Hacker News; BleepingComputer; Sysdig; watchTowr; US Department of Justice.

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.

For the past, I don't know, few years maybe, you and I have been sold this really beautiful, frictionless vision of the future. Right, the whole artificial intelligence as magic in a box pitch. Exactly. I mean, the pitch is always the same, right? You buy the super expensive corporate subscription, you plug it into your company's systems, and just boom. Instantly, your entire business is transformed. Yeah, you're supposed to get immediate efficiency, massive profit margins, and a workforce that suddenly just has superpowers. It's an incredibly seductive idea. Yeah. We all

want the easy button, especially when you're navigating these massive, confusing shifts in technology. Oh, absolutely. But, well, when we look at the diagnostic landscape from Kinsoft, specifically covering the first week of July 2026, we are presented with a massive reality check. A huge one. Yeah. We are looking at an industry -wide pivot. The tech giants are moving rapidly away from that plug -and -play myth, and they are colliding hard with the reality of how these tools actually... function or fail to function inside a real organization. So today's

deep dive is all about cutting through that hype. We're looking at a stack of fresh reports and data to see what actually works, what fails spectacularly, and where the real threats lie for you and your organization. Because the landscape is shifting under our feet as we speak. It really is. We are going to explore how major tech giants are fundamentally changing how they sell AI, the sudden rise of hyper -capable autonomous agents, and how these shifts are, well, completely rewriting

the cybersecurity battlefield. It's a lot to cover. Okay, let's unpack this. We have to start with the biggest and honestly the most surprising theme from the sources. Because the headline story this week isn't a new model release. Right. No flashy new gadgets. Yeah. It's the fact that the biggest AI providers on the planet are changing their entire business models right before our eyes. The shift is entirely structural. On July 2nd, Microsoft launched something called the Microsoft Frontier Company. And this isn't just like a small consulting wing, right? No, not at all. This is a $2 .5 billion operating business.

They staffed it with around 6 ,000 engineers and industry specialists. Wow. Yeah. And their explicit mandate is to embed directly inside customer organizations to build and run AI systems right there on site. That is wild. And just two days prior to that, so on June 30th, Amazon's AWS stood up a $1 billion Ford deployed engineering group. Right. The parachute pods. Exactly. The sources literally describe them as pods of engineers who parachute into a business for 45 days. And

their whole mission is to just build working technology, make sure the client is actually self -sufficient, and then extract themselves. And we should note, OpenAI and Anthropic launched very similar ventures back in May. So every major provider has suddenly realized that just handing you a software license and wishing you luck isn't working anymore? Not at all. And the why behind this is staggering. The sources point to some MIT research showing that... 95 % of enterprise generative AI pilots deliver zero measurable impact on the bottom line. 95%. I mean, that statistic should make any executive pause. Yeah.

Zero impact. It indicates that despite the immense capability of the underlying technology, nearly everyone is failing to convert it into. actual measurable business value. Right. So the vendors have looked at this massive failure rate and realized, hey, the bottleneck isn't the intelligence of the model. Because the models work beautifully in a vacuum. Exactly. The bottleneck is the messy, complicated reality of corporate implementation. To put this in perspective for you, imagine going

to a store and buying a massive state -of -the -art IKEA kitchen. Okay, I like this. You've got all the flat pack boxes. You've got the glossy catalog. You're incredibly excited to start cooking. But there's a problem. A huge problem. Ikea looks at their customer data and realizes that 95 % of the people buying this beautiful kitchen. They don't actually know how to use an Allen wrench. Right. And their home plumbing doesn't match the Swedish metric system. And I don't know, their drywall is just crumbling. Structural disaster. Exactly. So rather than letting their

expensive products sit in boxes in your living room, gathering dust and risking you calling them up and asking for a refund. IKEA realizes they have no choice. They have to step in. Yeah, they are suddenly forced to dispatch 6 ,000 master carpenters directly to your house just to make sure the kitchen actually gets built. What's fascinating here is how this entirely alters the relationship between you and the vendor. How so? Well, because corporate data is so deeply siloed and legacy workflows are so entrenched, change management is just incredibly difficult. So these tech giants are essentially taking over

the implementation phase entirely. Which sounds great on paper. It does, but there is a massive caveat for you to consider here. Having Microsoft or Amazon send their elite engineers to build your custom AI solutions, it's fast. It absolutely clears that immediate hurdle. But there's a catch. A big one. It deeply entrenches your dependence on their specific platforms. Oh, I see. You are effectively handing them the architectural blueprints to your entire operational infrastructure. Exactly. It really raises the stakes on the classic build

versus buy versus partner dilemma. Because if you partner to this degree, the lock -in is severe. So if your own internal AI pilots have quietly gone nowhere, you shouldn't feel bad. You are overwhelmingly the norm. You are the 95%. But you have to be incredibly clear -eyed about who these parachuting engineers actually work for at the end of the day. Right, because they aren't your employees. No. They are there to integrate your company so deeply into their proprietary ecosystem that leaving them later on becomes,

well, financially ruinous. They work for the platform's bottom line, not yours. Precisely. So this begs a massive question. If enterprise implementation at the top level is failing 95 % of the time and it requires a literal SWAT team of engineers to fix, how do everyday businesses actually afford to run AI at scale? That is the million -dollar question. And the sources point to the newly released mid -tier models as the missing link here. Okay, mid -tier, like Anthropic's new release. Yes. On June 30th, Anthropic released

Claude Sonnet 5. And they claim it performs very close to their top -end Opus model, but at an absolute fraction of the cost. We are talking... $2 per million input tokens and $10 per million output tokens. Right. And the terminology there is really important to clarify because the underlying mechanics dictate the economics of all this. Okay, break that down for us. Think of a token roughly as a piece of a word or maybe a syllable. And the reason there is such a stark difference in price between input and output comes down to computational effort. Got it. Input tokens

are what you feed the AI it is essentially reading. It costs Anthropic far less computing power to have their model scan a million tokens of your internal corporate documents than it does to have the AI actively generate or write a million new tokens from scratch. Which is the output. Exactly. Output takes much more compute. So if I'm understanding this, input is like handing the AI a textbook and telling it to read it. And output is asking the AI to write a completely original thesis based on that textbook. That's a great analogy. Writing takes way more brainpower,

so they charge you five times as much for it. Precisely the mechanism at play. But the critical detail with Sonnet 5 isn't just the drastic drop in the price of these tokens. It's the actual functionality. What can it do that's different? Well, this model is specifically tuned to run as an autonomous agent. It isn't just a chat interface sitting there waiting for your prompt. Right. It can plan multi -step tasks. It can navigate a live web browser or operate a command line terminal completely on its own to achieve a goal you set for it. Okay, here's where it

gets really interesting, though. And I kind of want to push back on the logic here for a second. Go for it. We just established that 95 % of corporate AI pilots are failing because corporate infrastructure and data is a complete mess. Right. So if the bottleneck is implementation and messy data, does making a cheaper mid -tier model that acts autonomously actually solve that implementation bottleneck? Or does this just make it cheaper for companies to fail at a much higher volume? That's a very fair point. But if we connect this

to the bigger picture, the mid -tier models actually solve a massive piece of the unit economics puzzle. How so? Well, the headline models, the massive expensive flagship AIs from OpenAI or Google, they get all the media attention. But the mid -tier is where the real daily work actually happens. Because it's cheaper. Yes. Think about it this way. If you priced up a business automation six months ago, say, having an AI independently open, review, cross -reference, and process thousands

of vendor invoices, the numbers likely didn't stag up at all. Because using the flagship model for that would be too expensive. Exactly. The compute cost of having a top -tier model do that repetitive background work would completely eat your entire profit margin. Right, because you can't have an AI doing thousands of mundane, everyday tasks if it costs you a dollar every single time it clicks a button. Exactly. But this new pricing structure completely changes the math for your business. Suddenly, processing those invoices autonomously costs pennies. It actually makes financial sense. Oh, I see. However,

utilizing an autonomous agent like that changes your risk profile entirely. Security -wise. Yes. An agent with access to a browser or a terminal needs to be treated like a brand new human employee. It requires strict guardrails, extensive activity logging, and crucially, least privilege access. Just to clarify that concept for the listener, least privilege means you treat the AI exactly like you would, say, a brand new summer intern. That's the perfect way to look at it. You give

the intern the keys to the specific filing cabinet they need for today's project, but you absolutely do not give them the master passcode to the server room so they can accidentally delete the entire company database. Yes. If you let an AI loose in your terminal with sweeping administrative rights, a single hallucination or just a misunderstood instruction could result in absolute catastrophe. Which is a terrifying thought. But, you know, if legitimate businesses are rushing to build infrastructure for these cheap agents to automate their accounting and HR departments, it tells

us something massive is shifting underneath the surface. The plumbing is being upgraded. Exactly. To power millions of new, cheap, autonomous agents, an unbelievable amount of capital is quietly flowing into the underlying plumbing of the AI world. We are seeing massive investments there. Yeah. On July 1st, a firm called Together AI, which essentially provides infrastructure to help businesses run open source AI cheaply, they raised $800 million. Wow. Yeah, that puts their

valuation at an astonishing $8 .3 billion. And just purely as a factual financial detail from the sources, that funding round was led by the venture arm of the Saudi oil giant. That capital influx is very much part of a broader trend of fortifying the foundation layer. Yeah. Bloomberg also recently reported that Meta is building a commercial cloud business of its own, internally called MetaCompute. Right, with the goal of renting out its spare AI computing power to outside customers, though we should explicitly clarify that this is an unconfirmed press report based on anonymous

sources. Yes, definitely. We are treating it purely as reported, not official. But the trend is clear. So what does this all mean for you? When I see oil giants and social media monoliths pouring billions into server farms and open source hosting frameworks, it feels exactly like the 1849 gold rush. The pickaxes and shovels. Exactly. Everyone remembers the miners out in the rivers panning for gold, hoping to strike it rich on a lucky find. But the people who built the real generational fortunes weren't the gold miners. No, they weren't. It was the people selling the

pickaxes, the people manufacturing the durable denim jeans, and the people building the railroads to transport the gold out of the hills. That is incredibly accurate. The models themselves, the chatbots, the autonomous agents, they are the miners. But the massive capital is flowing into the layer beneath the AI you actually touch. The railroads. Yes, it's going into the data centers, the compute power, and the hosting frameworks. And honestly, for you, the listener, this massive influx of capital going into the plumbing is

entirely positive. Because it drives down prices. Exactly. It means fierce competition down at the foundation layer. And over time, that competition results in better choices, more stability, and most importantly, much cheaper computing prices for anyone trying to run these systems. OK, so we've established that these autonomous AI agents are getting incredibly cheap. They're highly capable and they have this massive bedrock of. well -funded infrastructure supporting them. Which is great for innovation. That is fantastic

news if you are trying to automate your supply chain. But it terrifies me to think about what happens when malicious actors get their hands on that exact same cheap computing power. Yes. The implications for cybersecurity here are profound. We are no longer dealing with theoretical threats. The sources detail research from Sysdig, who documented a piece of malware they actually characterized as the first fully agentic ransomware. And they named it Jada Puffer. Right, Jada Puffer. And the key distinction here is that this wasn't, you know, a human hacker sitting in a dark room

using Chad GPT to help them write malicious code faster. No, not at all. This was an AI agent acting end to end completely on its own. That is so creepy. It really is. It broke in, it hunted through the network for secrets, and it executed the extortion autonomously. How did it even get in? It achieved this by exploiting an unpatched flaw in an open source AI development tool called Langflow. And once it breached the perimeter, the AI agent methodically harvested cloud keys, grabbed database credentials, and successfully

encrypted over 1 ,300 configuration items in a targeted system. But here's the specific detail that absolutely blew my mind. Right. During the attack, one of the AI's login attempts actually failed. It hit a roadblock. Right. It got rejected. But instead of stopping or crashing, the AI agent diagnosed the problem, rewrote its own approach, and had a working fix. running just 31 seconds later. This is the mechanism we really have to linger on because this machine speed adaptation changes the entire security paradigm. 31 seconds.

Think about it. When a traditional automated script hits a failed login, it usually just crashes and generates an error report for the human to review later. Right, it just gives up. But Jada Puffer parsed the server's error response, understood logically why it failed. Maybe the API endpoint had changed or the payload formatting was just slightly off and then it rewrote its own Python script to adjust to the new parameter. And just executed it again. Exactly. Wait, 31 seconds. If a human hacker hit a failed login. and got

some unexpected error code, they would still be taking a sip of their energy drink. They'd be cracking their knuckles, opening up a developer forum, trying to figure out what went wrong. It would take a human 10 minutes just to realize they had a typo in their code. It's true. Doesn't this mean the skill floor for launching a devastating targeted cyber attack just dropped straight into the basement? It really does. This raises an incredibly important question about the democratization of cyber threats. The skill floor has dropped to the literal cost of renting an AI. Exactly.

Sysdig's characterization of this being the first fully agentic ransomware might be their specific marketing framing, but the core lesson is undeniable. These autonomous agents can sweep up easy wins at a speed no human could ever match. They are constantly scanning the internet for exposed developer tools, unpatched software, default passwords. And they don't sleep. They don't sleep, they don't get distracted, and as we saw, they adapt to roadblocks in 31 seconds. So, how do you actually defend your business against a hyper

-fast, autonomous AI hacker that never sleeps? You'd think you need some sci -fi defense system. Right. You would naturally assume the answer is to go out and buy a million -dollar, next -generation AI defense shield to fight fire with fire. But surprisingly, the sources indicate the answer isn't more AI. It's not. No. The answer is fixing incredibly boring, traditional vulnerabilities and securing the human layer. Boring defenses. Exactly. The most advanced technology in the world doesn't matter at all if you leave the

architectural vulnerabilities exposed. Let's look at the boring defenses first. On June 30th, Citrix released a patch for a critical flaw in its Netscaler networking gear. It's tracked as CVE -2026 -8. And this is a big deal, right? A massive deal. Security experts are comparing the potential fallout to Citrix Bleed, which caused immense operational damage a couple of years ago. This new vulnerability is a pre -authentication memory leak bug. Okay, I want to make sure I understand how this actually works. A pre -authentication

memory leak. It's a mouthful. Yeah. So to understand this, imagine a bouncer standing at the front door of an exclusive VIP club. The bouncer is the authentication system. Okay. Normally, you hand him your ID, he checks his list, and he lets you in. But with this memory leak bug, instead of showing an ID, an attacker tricks the bouncer into dropping his clipboard. I see where you're going with this. And when the bouncer bends down to pick it up, he accidentally blurts out everything he is currently thinking about, which happened

to include the private VIP list. That is an excellent way to visualize it. When the device spills chunks of its own active memory, an attacker can just sift through that raw data and find active session tokens. The VIP list. Yes. Those session tokens are the VIP list. Once the attacker extracts a valid session token, they don't even need to show an ID or a password. They just walk right in. They bypass the login screen entirely. They bypass your multi -factor authentication, stepping right into the network as a fully authorized

user. Wow. And security researchers saw this specific net scaler vulnerability being exploited in the wild within 24 hours of the patch landing. They move fast. So if you run this gear facing the open Internet, the source makes it clear. This is a patch it today emergency, not a wait for the maintenance window next month kind of thing. And that is exactly what we mean by a boring defense. Diligent patch management isn't glamorous. No one is throwing a party for patching a server. Right. It doesn't make for an exciting keynote presentation. But it is the primary barrier

standing between your network and an automated AI agent looking for a fast, easy win. But the technology is only half the battle here, which brings us to the human element. Always the weakest link. Yeah. On July 1st, the U .S. Department of Justice announced the extradition of a 19 -year -old dual U .S. Esconian citizen from Finland. Right. The scattered spider case. Exactly. He is an alleged member of the notorious scattered spider hacking crew, and he's accused of helping breach a luxury retailer to demand about $8 million

in cryptocurrency. And we must note, these are allegations not yet proven in court. Of course. But the tactic itself is what we really need to focus on. Because Scattered Spider's main weapon isn't high -tech machine speed AI. No, it's not. It is just old -fashioned social engineering. They don't write complex code to break into the system. They manipulate humans to talk their way in. It's so simple. It is. They locate a target, gather some basic information, and then they literally just call the human IT help desk.

On the phone. On the phone. They pretend to be a legitimate employee who has, you know, been locked out of their account. And they politely ask the help desk to reset their access. When you put these two distinct threats side by side, it's just mind boggling. What do you mean? It's like installing a million dollar retinal scanner on the front door of your business, but leaving the back door completely off its hinges, which is ignoring the Netscaler flaw. Right. Or worse, having a thief just call the security guard on the phone. say they lost their badge, and smoothly talk the guard into handing over the master keys,

which is the scattered spider tactic. That's a perfect summary. The best defenses against both machine speed AI attacks and human social engineers rely on fundamental IT hygiene. The basics. Yes. You must patch your internet -facing systems immediately. You must kill default credentials. You cannot expose your raw development tools to the open internet. And what about the human side? Well, when it comes to defending against humans, the most advanced firewall in the world is useless if the human protocol is weak. You

have to empower your help desk. They need to be extensively trained and, frankly, given the structural authority to rigorously verify someone's identity before they ever reset a password or hand over network access. Even if the person on the phone is yelling at them. Especially then. Man. We have covered a massive amount of ground in this deep dive. Let's recap the journey really quick. Sounds good. We started with the startling reality that 95 % of corporate AI pilots are failing to deliver a measurable return, which is forcing tech giants like Microsoft and Amazon

to literally send armies of engineers into businesses just to make the technology function within messy corporate data silos. The IKEA carpenters. Exactly. Then we saw the rise of cheap mid -tier... AI models like CloudSonic 5, making autonomous agents financially viable for everyday business automations. Pennies per task. Right. But we also saw the dark side of that cheap computing power, with the rise of Jada Puffer and autonomous AI agents capable of hacking a system, diagnosing a roadblock, and rewriting their own code in just 31 seconds.

Still hard to believe. It is. And finally, we learned that the best defense against all of this isn't necessarily more AI. It is diligently patching your systems today and ensuring your help desk can't be fast -talked by a 19 -year -old hacker. It is a rapidly evolving, highly complex landscape, but the fundamentals of how we protect our organizations remain surprisingly constant. They really do. However, I want to leave you with a final thought to ponder, something that connects all the diverse dots we've discussed today. I'm ready. Lay it on us. Today, we talked

about AI agents acting autonomously to hack systems. And we talked about human hackers using social engineering to trick human help desks. Right. Two separate threats. But look at where these two distinct trend lines intersect. Okay. As AI agents like CloudSonic 5 get incredibly cheap, highly capable, and fully autonomous. What happens when the entity calling your help desk to social engineer a password reset isn't a 19 -year -old kid reading a script? Wait. What if it's an autonomous conversational AI agent that has been trained

to sound exactly like your CEO? Oh, wow. That is terrifying. How will your human employees verify if the frantic voice on the other end of the phone demanding urgent access to the network is actually a co -worker in distress or just a 31 -second script running on a mid -tier model? They wouldn't know. That is the intersection we were rushing toward, and it is a reality businesses need to prepare for today. That brings us right back to where we started. We all wanted the plug -and -play magic box that would make everything easier. And we didn't get it. No. What we got

was a landscape where you have to meticulously guard your internal data, patch your systems daily, and prepare for a future where you can't even trust the voice on the other end of the phone. It's a brand new world. It really is. Yeah. And the easy button. definitely does not exist. Thank you for joining us on this deep dive. Stay curious, stay patched, and keep exploring.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Iris Ohyama - A Suspicious File in the Cloud, 7.7 Million Customers at Risk, and an Online Store That Went Dark

    Nov 19, 202620 minAudio

    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...

    Credit the people on this episode

  2. Who Gives A Crap - A Marketing Supplier Breached, Six Days to Tell the Client, and Customer Data That Didn't Need to Be There

    Nov 17, 202619 minAudio

    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...

    Credit the people on this episode

  3. Last Week in Tech - An Outlook Zero-Day That Leaks Your Password Hash, Chrome's Eighth Exploited Bug, and Nvidia Without Jensen Huang

    Nov 15, 202621 minAudio

    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...

    Credit the people on this episode

  4. Chainalysis - From "Non-Sensitive" to "All Users": How a Breach Notice Changed in Five Days

    Nov 12, 202619 minAudio

    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...

    Credit the people on this episode

  5. NSW Health - A Dental Voucher Website, Four Gigabytes Taken, and the Providers Whose Details Were Exposed

    Nov 10, 202623 minAudio

    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;...

    Credit the people on this episode