Last Week in Tech – Google's Agentic Gemini, the Coding-Agent Gold Rush, and CISA's Own Key Leak
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: Google goes all-in on agents at I/O 2026 with Gemini 3.5 and a new any-input video model, Gemini Omni. Three rival coding agents launch in three days as the price of frontier coding power keeps falling. SpaceX's S-1 lifts the lid on xAI's finances, and Anthropic teams up with the Gates Foundation on a $200M global-development push. On security: a CISA contractor accidentally exposes AWS GovCloud keys on a public GitHub repo, Canadian police arrest the alleged operator of a massive IoT botnet, and a Vietnamese government network is breached. Building or buying AI tools and not sure how to keep your secrets out of public repos? Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: Tech Startups; Medium (David Akpovi); Digital Applied; The Hacker News; 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.
You know, usually when you hire an assistant, there's this expectation of supervision. Right. Yeah. Like you have to manage them. Exactly. It's sort of like being a driving instructor. You're sitting right there in the passenger seat. You've got your own set of brakes. And, you know, when the student driver veers a little too close to the curb. You just reach over and correct the wheel. Right. You gently correct the steering wheel. Because ultimately, you're the one in control. You are fundamentally in control of the vehicle. Yeah. The machine. Or the assistant assists, but the human decides. It's a very comforting
safety net. Because the ultimate responsibility rests in your hands. But if you look at the tech landscape right now, it's like suddenly that passenger side brake pedal is just gone. Completely gone. We're looking at a technological shift where you aren't just, you know, letting go of the wheel. You're basically standing on the sidewalk watching the car drive itself across the country. Right. And it's negotiating for its own gas, booking its own hotel rooms along the way. Yes. And that's exactly what we're getting into today. Welcome to today's deep dive. Our entire mission
today is unpacking this massive paradigm shift. We are moving entirely out of the assistant era and diving headfirst into the agentic era. It is a massive transition. It really is. And we're pulling all this from a really fantastic source today. It's the Kinsoft Tech Talks update from May 25, 2026. So just a concise, high -level tech industry recap that cuts through all the noise. And the reason you, the listener, really need to care about this specific update is because this transition to the agentic era, it fundamentally
alters how businesses have to think about automation. Oh, absolutely. It brings these multibillion dollar financial ripple effects and honestly, some security implications that are just terrifying. Yeah, we're going to get to those later. And they are wild. They really are. And whether you're just a casual tech observer or you're running a business, you need to understand this. This is the new reality. OK, let's unpack this. We have this specific week. in May 2026 that serves as like a tipping point. And it starts with Google I .O. Right. Their big developer conference. Yeah. They stood on stage and essentially went
all in on this agentic era. They announced Gemini 3 .5, which is geared specifically for agents and coding. But the one that really grabbed me was Gemini Omni. Oh, Gemini Omni is a completely different beast. That model represents a massive architectural leap. How so? Well, the distinction between older systems and Omni is really critical to understand. Like earlier AI models were essentially these separate brains that were just clumsily taped together. Right, like different departments in a company. Exactly. A text model processed
the text, an image model processed pixels, and an audio model handled the sound waves. So when you asked a complex question that involved... looking at a photo. They had to hand the information back and forth to each other. Right. It's kind of like a committee trying to describe a movie scene to each other over the phone, like it's lost in translation. A highly inefficient committee and with incredible latency too. But Gemini Omni changes the underlying architecture by being natively multimodal. Meaning it's all one system.
Yes. It doesn't translate images into text to understand them. It ingests the video, the audio, text and images all through the exact same neural network simultaneously. Oh, wow. So it's processing reality a lot closer to how we do, right? Like human sensory integration. Exactly. Which allows it to instantly understand deep context. And then it can generate high quality video or audio right out the other side without all that lag. That's incredible. But, you know, Google wasn't acting in a vacuum here. The pace described in
this Kinsoft update is just staggering. It really was a crazy week. I mean, within a span of three days, three entirely separate frontier -grade coding agents dropped. Three days. Yeah. Cursor released Composer 2 .5. Anthropic dropped code with Claude. And Alibaba launched Quinn 3 .7 Max. Three massive releases in basically a weekend. And you really have to notice the terminology shift there. We're moving from assistants to agents. That isn't just marketing spin, I promise you. No, it means something very different. It
represents a fundamental change in how the software actually operates. I love the intern versus senior manager analogy for this. The assistant era we're leaving behind. It was like having a very eager but very fast intern. Right. They mean well, but you have to check everything. Exactly. You could ask for a block of code, but you had to read every single line, spot the hallucinated variables, fix the syntax errors. You were the manager constantly correcting the intern's work. While an agent, on the other hand, operates on a continuous autonomous reasoning loop. Autonomous
being the key word there. Exactly. You give the agent a high -level goal. Say you tell it, uh... build a secure login page that connects to my customer database. Okay, a standard task. Right. The agent plans the architecture, writes the code, but then it actually spins up a virtual environment and runs the code itself. Wait, it tests its own work? Yes. When the code inevitably fails, the agent reads the terminal error message, understands why the database connection dropped, rewrites the problematic function, and then tests
it again. So it's doing the debugging in the dark, basically, without bother you at all. It iterates completely on its own until the goal is achieved. And what's fascinating here is how this completely upends the economics of software creation. Oh, totally. Because the cost of frontier -grade coding help is just falling off a cliff right now. Because you have Google, Anthropic, Cursor and Alibaba caught in this vicious price war. They are trying to outcompete each other week by week. And it's a price war that fundamentally
alters the value proposition of the entire tech sector. I mean, think about a mid -level SaaS startup software as a service. Let's say they charge $50 a month for a specialized inventory management tool. Historically, businesses paid that subscription because building their own custom tool would require, you know, hiring a development team for six months. Right. And spending $100 ,000 easily. Exactly. But if a business owner can just open their laptop, tell an AI agent, build me an inventory tracker tailored
to my exact warehouse layout and workflow. And the agent just executes that entire reasoning loop for, what, $3 in compute costs? Yeah, exactly. Why would they ever pay for a generic SaaS subscription again? They wouldn't. The barrier to creating complex, bespoke applications is dropping to almost zero. So the dynamic between the people who build software and those who buy it is just completely compressing. The intrinsic value is no longer in writing the code itself. The value shifts entirely to the massive infrastructure running these agents and, of course, the proprietary
data you feed them. So what does this all mean for the companies actually footing the bill? Because, you know, having an AI senior manager constantly running these autonomous coding loops, testing and retesting, that requires a ridiculous amount of compute power. An unbelievable amount. You can't just run these frontier models on a standard laptop, which really explains the terrifying financial numbers we are seeing behind the scenes. Yeah, we got some rare financial sunlight on that front recently, thanks to a SpaceX S1 filing, oddly enough. Right, because of the Elon Musk
connection. Exactly. Because he intertwines his company's regulatory filings for one often exposed data for another. And this one accidentally contained the financials for XAI for the year 2025. And the numbers are just wild. I mean, XAI brought in about $3 .2 billion in revenue. Which is huge. Earning $3 .2 billion in a single year for a relatively new company is phenomenal. Against that revenue, they posted a staggering $6 .4 billion operating loss. Yeah, a massive burn
rate. Wait, stop. I have to push back here. Because in any normal industry, burning twice what you make isn't a business strategy. It's a bankruptcy hearing. It really is. How do you lose over $3 billion and convince investors you're doing a great job? Is this actually a sustainable business model? Well, it looks like financial ruin if you view it through the lens of a traditional software company. Right, where your expenses are just servers and payroll. Exactly. But AI infrastructure requires a completely different mental model. That $6 .4 billion isn't going to, like, marketing campaigns or bloated executive
salaries. Either. Where is it going? It is going toward a physical manufacturing race. We are talking about purchasing hundreds of thousands of specialized liquid -cooled microchips. Oh right, the hardware. Yeah, they're constructing massive data centers from concrete and steel and trying to secure the energy contracts to power facilities that literally consume as much electricity as a small city. This is a capital expenditure. They're treating AI compute like they're... I don't know, laying transatlantic
fiber optic cables or building a national electrical grid. That's a perfect way to look at it. They are funding it ahead of profits because they believe the winner takes all dynamic in AI is absolute. If you aren't first, you're last. Whoever builds the most capable physical infrastructure today will essentially own the cognitive autonomy of tomorrow. Wow. If a company pauses its capital expenditure to become profitable right now, their models just fall behind the competition and they become entirely irrelevant in six months. That is a brutally high stakes game of poker. Extremely
high stakes. But what really jumped out at me from the Kinsoft update is the duality happening simultaneously. Like right next to that hypercapitalist reality, we see Anthropic and the Gates Foundation launching a $200 million four -year partnership. Yes, a highly targeted effort to build AI tools specifically for health, education, and agriculture in underserved regions. It's just such a fascinating contrast. You have these multi -billion dollar bloodbaths between tech giants trying to dominate the future economy. Juxtaposed with this heavily
-sunded philanthropic initiative to deploy that exact same AI to help, like... A farmer in a developing nation optimized their crop yield today. It really captures the central tension of the AI boom beautifully. The foundational models are so incredibly resource intensive that they demand unprecedented corporate capital just to create. Right, the $6 .4 billion losses. Exactly. Yet their potential to solve fundamental human problems is so immediate that major philanthropic organizations are just bypassing traditional aid models entirely. They want to deploy AI directly
into the hands of the people who need it most. But, and here's the catch. As powerful as these tools are, whether they are writing custom software autonomously or optimizing global agriculture, they operate on networks secured by humans. Yes. And humans, unfortunately, make mistakes. We do. And this raises a really important question. What happens when you hand the keys to an incredibly powerful autonomous AI agent, but the lock on the door was installed by a human who just forgot to tighten the screws? Which brings us perfectly
to the security breakdown in the Kinsoft update. And here's where it gets really interesting because the cautionary tale they share is just dripping with irony. We are talking about CISA. America's cyber defense agency. The people whose literal mandate is to secure the nation's digital infrastructure. The absolute... ATECs of security professionals. You don't get higher than CISA. And yet, a contractor for CISA left a public GitHub repository exposing AWS GovCloud access keys. Just sitting there
on the open internet over a weekend. It's almost hard to believe. To visualize this for you listening, a GitHub repo isn't just a digital bulletin board. It is the architectural blueprint of an application where developers collaborate and store their code. Right. So leaving GovCloud keys in a public repo is like publishing the blueprints to the Federal Reserve online and accidentally leaving the master vault code scribbled in the margins for literally anyone to read. It is a devastating lapse in operational security, and the source
material uses this to make a highly sobering point for anyone running a business right now. Yeah, if it can happen to them. Exactly. If a contractor working for the nation's premier cyber agency can leak high -level cloud keys into a public environment, everyday developers working at a mid -sized company are perfectly capable of making the exact same error. Especially if those everyday developers are suddenly using the AI coding agents we just talked about to generate... software 10 times faster. If we connect this to the bigger picture, this is exactly why
the rise of AI agents makes stringent automated security protocols entirely non -optional now. You can't just rely on the honor system anymore. No, because the velocity of development has fundamentally changed. If a human is generating and deploying code at lightning speed because an agent is doing the heavy lifting, they are also capable of deploying catastrophic security vulnerabilities at lightning speed. Right. The mistakes scale with the productivity. Exactly. When a developer is managing a massive volume of AI -generated code, human oversight
just gets stretched too thin. You can't manually catch a stray string of text containing a hard -coded password. Exactly why Kinsoft lays out three mandatory security rules for this new era. And the first one is secret scanning on all code. And we really need to look at the mechanics of how that works in practice. Secret scanning is an automated pipeline step. Meaning it happens in the background? Yes. Before a developer's code is allowed to be published or merged into a shared repository, an automated tool actively reads every single line. It's mathematically
searching for patterns that look like passwords, API tokens, or access keys like those AWS GovCloud keys that were leaked. So it intercepts the blueprint and scrubs the margins before it ever reaches the public domain. Exactly. It takes the human error out of the equation. Okay. And the second rule they emphasize is short -lived credentials. Think of short -lived credentials like a digital hotel key card rather than a metal house key. Oh, that's a great way to put it. Right, because instead of issuing a developer or an AI agent a permanent password to access a sensitive database,
the system generates a temporary mathematical token, and that token expires automatically after a set period, say, four hours. So even if that token... Does leak onto the public internet because someone bypassed the secret scanner somehow? The window of opportunity for an attacker is incredibly narrow. The key simply turns to dust. I love that. And the final rule is least privilege access. This is the principle of containment. A developer or an AI agent acting on their behalf
should only have the exact system permissions required to do their specific task and absolutely nothing more. You don't give them the keys to the kingdom. Exactly. If an intern is hired to restock the lobby fridge, you don't give them the master keys to the CEO's office. By enforcing least privilege access digitally, if an account is compromised, the attacker is trapped in that specific room. The rest of the network remains secure. Precisely. Which is so critical right now because the threat landscape Kintoff describes is relentlessly hostile. I mean, they highlight
a 23 -year -old in Canada who was just arrested for running a massive Internet of Things botnet. Yeah, that was wild. And it was responsible for huge denial of service attacks. And Internet of Things, or IoT. Botnet relies on hijacking poorly secured smart devices. We're talking about internet -connected security cameras, smart refrigerators, or even office thermostats. Stuff you don't even think about having a password. Right. The attacker exploits basic vulnerabilities to compromise hundreds of thousands of these low -security devices, essentially turning them into a zombie
army. And then what do they do with them? They command all of those devices to send junk data to a single target server simultaneously. It just overwhelms its capacity and knocks it completely offline. It's terrifying. And in that exact same week, a Vietnamese government network was breached. I mean, the attacks are constant. They're highly sophisticated. And as we saw with CISA. All it takes is one exhausted contractor making a human error to let them in. The agentic era accelerates both sides of this equation, unfortunately. Right.
We are gaining unprecedented capabilities with multimodal models like Gemini Omni that synthesize reality and autonomous coding agents that build complex software loops from scratch. But trusting that automation requires an ironclad... automated foundation of security. We can't have one without the other. Exactly. We are rapidly reaching a point where the capability of the technology is no longer the bottleneck. Our ability to safely contain and manage it is the true limit. So let's just step back and look at the journey we've taken through the source material today. I mean,
we started by exploring this massive paradigm shift. Moving from the assistants to the agents. Right. Moving from AI assistants you have to constantly supervise to... AI agents that autonomously execute complex reasoning loops. And that's all driven by breakthroughs in multimodality and fierce industry competition. And then we examined the staggering financial reality of funding that shift. We looked at the billions being burned on physical compute infrastructure to win this AI arms race, while simultaneously seeing hundreds of millions deployed to solve immediate global
challenges in health and agriculture. And finally, we grounded all of that soaring potential in the unforgiving reality of cybersecurity. recognizing that as we let AI accelerate our development cycles, we have to implement automated fail -safes like secret scanning and temporary credentials. Because human error scales just as fast as human productivity. Exactly. For you listening to this deep dive, the application is direct. Look at your own workflows or your own business. The tools to automate your daily friction are getting cheaper and more powerful by the hour. But you
really must ask yourself, are your security practices and oversight mechanisms evolving at that exact same speed? Because if they aren't, you are building a skyscraper on a foundation of sand. And we really are. And that leaves us with one final... lingering question to ponder as we wrap up today. As these coding agents become cheaper and flawlessly capable, it is totally inevitable that we will eventually task AI agents with writing the security protocols to lock down our cloud systems. But in a world where AI is actively defending the
digital vault against other AI attackers, will we finally trust the machines completely with our safety? Or will good old -fashioned human error, a tired developer, a misconfigured setting, or a forgotten GitHub repo remain the ultimate unpatchable vulnerability that leaves the front door wide open? Something to chew on. Stay curious, keep digging, and we will 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;...