Last Week in Tech – AI Writes Its First Zero-Day, OpenAI's Deployment Play, and the Canvas Mega-Breach
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 on the week that was, with a security lens. This week: researchers document the first known case of attackers using AI to develop a working zero-day exploit. OpenAI launches a $4B Deployment Company and buys consultancy Tomoro to get AI into enterprises faster, while SoftBank pours money into AI data centres and batteries. On the breach desk: ShinyHunters complete what's now called the largest education-sector breach on record, defacing Canvas login portals at around 330 institutions - including Harvard and Princeton - before Instructure paid up. Plus Washington turns up the heat on AI, with the Pentagon flagging a major AI lab as a supply-chain risk. Worried about AI-accelerated attacks on your own systems? Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: Tech Startups; Coaio; cm-alliance; SharkStriker; TechCrunch.
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 a burglar, right? Like a guy who doesn't just learn how to pick your front door lock. Right. But instead, they bring a machine to your porch that instantly invents a completely new state of physics just to walk right through the wood. Which is terrifying. Yeah, it really is. And for a long time, that kind of... on the fly, impossible to predict threat was, well, purely theoretical in the cybersecurity world. Mostly just white papers and lab tests. Exactly. But last week that actually stopped being science fiction. Attackers have officially used A .I.
to generate a working zero day exploit out in the wild. It's a huge milestone. Unfortunate, but huge. Completely. So welcome to today's Deep Dive. We are looking at a really pivotal update from Kinsoft Tech Talks. And this is dated May 18th, 2026. Glad to be here to unpack it with you. The overarching theme of the intelligence we're diving into today is this really jarring transition of AI. It has officially moved out of the theoretical research labs and is now violently colliding with real world cyber warfare. Yeah.
And corporate enterprise architecture, physical power grids, aggressive government regulation. It's hitting everything. It really is. So to you, our listener. OK, let's unpack this because the stack of sources we have today. Gives us a remarkably complete picture of that collision. And it is a profound operational shift we are talking about here. Right. We have documentation on those first AI generated zero day exploits, a massive record breaking breach in the education sector, and these billion dollar deployments
that are actually altering physical power distribution. Which sounds like a sci -fi movie, I know. It totally does. But the actual lessons for you, the listener, are surprisingly grounded. It all comes back to the basics of security and critical thinking. We spent the last few years discussing AI as a capability, right? Like what it might do, how smart it could get. Yeah, the potential of it all. Exactly. But now we're looking at AI as an infrastructure reality. And when you transition from a capability to an infrastructure, you suddenly inherit all the messy physical and
economic realities of the real world. Right, the plumbing of it all. So let's start right at the bleeding edge of the threat landscape. According to the Kinsoft update, researchers have documented what is believed to be the first real -world case of attackers leveraging an AI model to write a functional zero -day. Which is just wild to see actually happen. It is. I mean, that means they exploited a vulnerability that the software vendor had absolutely no idea existed, meaning there was zero time to patch
it. Like I said at the start, it's like a burglar who doesn't just buy a lockpick, but uses a machine that instantly invents a brand new type of lockpick the second they walk up to your front door. That's a great way to put it. And what's fascinating here is that the response shouldn't be panic. Really? Because it sounds pretty panic -inducing. I know, I know. But we need to look closely at the mechanism of how this actually happens because it alters the economics of cyberattacks. Historically, uncovering a zero -day required immense human capital. Right, like rooms full of people staring
at code. Yeah, highly specialized security researchers. They'd use techniques like fuzzing. you know, throwing random data at a program to see if it crashes or painstakingly reverse engineering compiled binaries. Just looking for a needle in a haystack. Exactly. They would spend months looking for a microscopic flaw in how an application handles memory, hoping to find some obscure execution path that leads to a buffer overflow. It was a very manual artisanal process. So it was constrained by human working memory, basically. Yeah. And
time. A human can only hold so many variables in their head while tracing a logic path through, I don't know, millions of lines of C++ code. Precisely. But an advanced AI model doesn't face those cognitive bottlenecks at all. It can ingest entire code bases simultaneously. Wow. Just all at once. Yeah. It uses graph neural networks to map every single execution path, every dependency, every data flow. It recognizes architectural patterns and can spot an uninitialized variable that human reviewers just gloss right over. Or
even traditional automated scanning tools, right? Exactly. The AI accelerates the vulnerability discovery timeline from months of manual labor to literally minutes of compute time. Minutes. That is insane. So what does this all mean for the defense? If an attacker has a machine that can generate a brand new unseen exploit on demand, how do you defend an enterprise network? You can't patch a vulnerability that nobody on the planet knew existed five seconds ago. And this is where we run into a really fascinating paradox.
The ultimate defense against an AI that can think a million times faster than a human isn't a faster, smarter AI. Oh, wait, really? It's not? No, it's actually static, dumb architecture. Because the speed of attacks is scaling up so fast, the defense relies on the absolute fundamentals of system design. Fast patching, strict network segmentation, and the principle of least privilege. Okay, let's break down the mechanics of that. If the AI lockpick gets through the front door using some novel technique, how does static architecture stop it? By ensuring the hallway leads absolutely
nowhere. Ah, I see. Even the most sophisticated zero -day exploit usually serves one primary function, and that's initial access. Once the malicious code executes, it inherits the permissions of whatever application or user it just compromised. Right. So if you have implemented zero -trust architecture and strict least privilege, That compromised user account has no administrative rights. It cannot access the central database, it can't alter system registries, and it can't move laterally to other servers. So the AI might
brilliantly engineer a way to crash an app and execute code, but then it finds itself trapped in a digital padded room. Exactly. It has no privileges to escalate, no network routes to traverse. A closed door that's mathematically enforced by a permission policy can't be outsmarted by a generative model. That makes a lot of sense. The mechanism of entry might be brand new and highly intelligent, but the behavior required to actually do damage, like stealing proprietary data, requires structural access. Right, and if you limit the access, you contain the blast
radius. Good monitoring detects the abnormal behavior in that hallway, regardless of the exotic zero -day that got them inside. Okay, but looking at the Kinsoft briefing, that reliance on foundational architecture brings up a glaring contradiction. How so? Well, if AI zero days are the terrifying new frontier, why are we still seeing record -breaking takedowns from standard ransomware attacks on vendors? Are we distracted by the shiny new AI object here? That is the million -dollar question. While we are analyzing the
threat of AI zero days, you're right. The group Shiny Hunters just executed the largest education sector breach on record. And they didn't use an AI zero day to do it at all. No. They used the most traditional leverage point in the modern enterprise, the third -party supply chain. The details on this are just staggering. Shiny Hunters targeted the edtech giant Instructure twice in a two -week period. They compromised the Canvas learning management system, basically defacing the login portals for roughly 330 distinct instances. Including Harvard and Princeton. We are talking a global infrastructure here. Yeah. And this
connects back to our previous deep dive, right? The Australian Canvas episode we did a while back. It just shows the global scale of this problem. They knocked the platform offline right in the middle of final exams. The timing is a masterclass in extortion economics, honestly. Oh, for sure. And Instructure ultimately paid the ransom just one day before the attacker's deadline. Right. By hitting the platform during finals, the attackers maximize operational pain and psychological pressure. But the truly critical lesson for you here is the architectural vulnerability
of the trust relationship. Your software vendor's security is your security. Right, because Harvard and Princeton have incredible internal security. They have massive IT budgets, advanced firewalls, excellent lease privilege architectures internally. But all that gets bypassed. Exactly. Why spend resources having an AI invent a brand new zero -day exploit to tunnel through Harvard's firewall when you could just compromise an API key from an edtech vendor that Harvard already whitelists. And if we connect this to the bigger picture,
it explains exactly how the modern enterprise network actually functions. A university or any Fortune 500 company cannot operate as an isolated island anymore. Right, you have to use vendors. You rely on cloud -based vendors for human resources, communication, and in this case, learning management. To make that seamless for users, the organization integrates the vendor using single sign -on protocols like SAML or OOOTH. up trusted network tunnels and exchanging API tokens. So the university
is essentially handing the vendor a master key to a specific section of their internal network. Trusting that the vendor's security is airtight. And that trust is the single point of failure. Shiny hunters didn't need to breach 330 heavily defended university networks. They only needed to compromise the central infrastructure of Instructure. Exactly. Once inside the vendor, the attackers leveraged those existing trusted API webhooks and SSO integrations to cascade the attack down to the clients. Because from the university's perspective, the malicious traffic is coming from a trusted, authenticated domain. The firewall
just lets it right in. It's a massive force multiplier for the hackers. No matter how good your internal AI defenses, relying on an external platform creates a massive bottleneck of risk. So if you're running a business, the reality is that your internal perimeter doesn't stop at your own servers anymore. Not at all. And that exact vulnerability, the reality that embedding a third -party vendor creates a structural risk, is what makes the next phase of the AI transition so incredibly volatile. Because the industry is undergoing
a massive pivot right now, shifting away from just training AI in a vacuum to aggressively deploying it into corporate supply chain. Right. OpenAI recently launched the OpenAI Deployment Company, and they backed it with over $4 billion in funding. They also acquired a consultancy called Tomorrow, bringing in like 150 deployment specialists. The acquisition of a consultancy is the defining signal of where the market is. For years, the objective was purely computational, right? Like, who can train the model with the most parameters? Who has the most advanced neural architecture? Yeah. But an incredibly smart raw
language model sitting on a server in California does not inherently generate revenue for a logistics company in Chicago. Or a hospital in London. It has to be integrated. And here's where it gets really interesting. The AI industry is transitioning from the lab scientists mixing the formula to the plumbers and salespeople trying to install the pipes in your house. That is a perfect analogy. But it feels like a massive identity crisis for an organization that set out to build artificial general intelligence. Why the sudden pivot to
the unglamorous world of corporate IT consulting? Because the bottleneck to AI dominance is no longer just algorithmic intelligence. It is corporate adoption. To make an AI useful to an enterprise, it has to securely interact with the company's proprietary data. You have to build complex retrieval augmented generation or RH pipelines. You have to map the company's data lakes, convert documents into vector embeddings, integrate the AI into their internal APIs. So it can actually do things like draft emails or query inventory databases.
The plumbing, exactly. And mapping corporate data lakes requires people. It requires deployment specialists to get their hands dirty in the client's internal architecture. Which explains the 150 consultants. OpenAI is spending $4 billion because they are racing to become... your most deeply embedded vendor. They want their models woven into the very fabric of your daily operations. But based on everything we just discussed with the Canvas breach, integrating a vendor that deeply insecutes is exactly how you expose your
company to a cascading supply chain attack. Exactly. If open AI or any major AI provider becomes the central nervous system for thousands of corporations, they become the ultimate target for state -sponsored hackers. Which is why enterprise leadership has to approach AI... deployment with extreme contractual defense. If a vendor offers to put AI into production for you, you welcome it, sure, but you must legally mandate data handling and security in the contract from day one. Exactly. You define exactly what data the model is permitted to ingest. You mandate
strict data retention policies to ensure your proprietary prompts aren't stored on their servers. And enforce that same static zero trust architecture we discussed earlier against the AI vendor itself. So you restrict the AI's API access. Right. So even if the vendor's central infrastructure is compromised, the attackers can't use the AI's connection to pillage your internal databases. That makes sense. But the friction of this deployment rush isn't just happening in software architecture, is it? It's hitting a hard physical limit. Yes,
the physical reality is catching up. The Kinzhoff briefing highlights that SoftBank is pouring massive capital into the physical side of this boom, standing up new ventures for AI data center hardware and, critically, next -generation batteries. To feed the power -hungry models. Let's explore the mechanics of that. Why is a massive telecom and investment firm suddenly obsessed with municipal power grids and battery technology just to run software? Because of the computational physics of how these models operate, when a user queries a large language model, it isn't just retrieving
a static file like a traditional web search. Right. It's generating it. Yes. It performs billions of complex matrix multiplications across thousands of GPUs simultaneously just to calculate the statistical probability of the next word. And GPUs, especially the advanced ones used for AI, draw an immense amount of wattage. We are talking about power requirements that are forcing companies to rethink global infrastructure. A single large -scale AI data center can demand gigawatts of electricity. That's the equivalent consumption
of a small city. The current physical infrastructure simply cannot support the widespread corporate deployment these AI labs are pushing for. They need advanced liquid cooling systems to prevent the silicon from melting and hyper -stable power supplies. A momentary dip in municipal power could disrupt training runs that cost, what, tens of millions of dollars? Yeah, exactly. So SoftBank realizes that the company that controls the energy storage, the advanced batteries that can buffer grid fluctuations and guarantee uninterrupted power, controls the actual choke point of the
industry. The software race has fundamentally become an infrastructure race. It really has. And when you're building technology that requires municipal -level power grids, maps the proprietary data of Fortune 500 companies, and holds the API keys to global supply chains, well, you stop being just a tech company. Right. You become a critical geopolitical asset, which explains the final major shift detailed in the Kinsoft briefing. Governments are stepping in. The era of unregulated shadow operation is closing. When
technology reaches the scale of critical infrastructure, it invariably attracts the heavy hand of the state. Yeah, the briefing notes that the political climate in Washington is really sharpening. Senators have officially floated the creation of a federal AI commission. And simultaneously, the Pentagon has reportedly labeled a major AI lab a supply chain risk, ordering agencies to cut ties and rip out their integrations. Now, we should... Be clear to you listening. We aren't endorsing the politics behind any of these moves. We're strictly reporting on the government's actions to understand the operational fallout. Absolutely.
We're just looking at the reality on the ground. But the whiplash for a normal business here is severe. So what does this all mean? How does a regular business navigate this whiplash of open AI aggressively pushing corporate deployment while the Pentagon is simultaneously blacklisting AI labs as supply chain risks? This raises an important question, and it really synthesizes this entire deep dive. The convergence of vendor reliance, physical infrastructure limits, and government regulation means AI is no longer just
an IT issue. Right. It is a core geopolitical and operational risk. If you build your entire company's workflow around a specific AI model, you inherit that vendor's geopolitical baggage. That is a terrifying thought. Because if the Pentagon declares your primary AI provider a supply chain risk overnight, and federal regulations mandate you sever ties, your business operations could instantly grind to a halt. Completely. Or if their local power grid fails because the infrastructure wasn't upgraded, your customer service pipelines go dark. If they suffer an
SSO breach, your intellectual property is exposed. So you really have to treat an AI provider with the exact same scrutiny and skepticism that a manufacturer applies to a factory producing critical aerospace components. Yes. You diversify your vendors. You maintain rigid isolation between your core data and their models. And you constantly monitor the regulatory weather report, I guess, to anticipate when your vendor might suddenly be blacklisted. Exactly. And that brings us full circle to the ultimate defense strategy. It does.
The sign -off from the Kinsoft brief is incredibly fitting for everything we've just mapped out today. They said, stay patched, stay skeptical. It sounds overly simple given the sheer complexity of gigawatt data centers and AI zero days, but the mechanics prove it out. They really do. The foundation of security doesn't change just because the adversary's tools get faster. Against a generative model hunting for zero days, against ransomware groups leveraging trusted API... tunnels, against the geopolitical instability of massive deployment vendors, your only mathematically sound defense
is mastering those basics. Strict network segmentation, absolute least privilege, ironclad data retention contracts, and relentless monitoring of your internal traffic. It is the paradox of modern tech. To survive the future, you have to be obsessive about the static architecture of the past. It's the most critical operational audit. any organization will conduct this decade. So to you listening, whether you are prepping for a meeting or just trying not to get overwhelmed by all this, the ultimate defense against both AI zero days and
massive vendor breaches is mastering the basics of monitoring and contracting. And if you're wondering if your patching can keep up, you can find more resources over at www .kinsoft .com .au. And as we close out this deep dive, I want to leave you with one final lingering question to ponder on your own. Let's hear it. We've seen the devastating cascading effects of relying on... massive third -party vendors during the Canvas breach. And we are seeing governments actively blacklist central AI labs due to supply chain risks. So if governments and massive corporations
are realizing that third -party AI labs are a dire supply chain risk, how long until organizations are forced to build smaller, localized, offline AI models just to ensure their own survival when the internet or their primary vendor inevitably goes down? Wow. A private intelligence entirely disconnected. Protected from the grid, protected by a closed door. That completely changes how we think about the future of this technology. It really does. Thanks for joining us on this deep dive. Stay patched. Stay skeptical.
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;...