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 – Claude Opus 5, a Patchless Java Zero-Day, and EY's Support-Desk Breach

Aug 2, 202620 minAudio

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Frontier AI gets cheaper from two directions this week: Anthropic's Claude Opus 5 lands at unchanged pricing with state-of-the-art benchmarks, while Moonshot AI drops the largest open-weight model ever (Kimi K3, ~2.8T parameters) - big implications for data-sovereign self-hosting. Google ships Gemini 3.6 Flash while reports say its flagship keeps slipping. In security: the FastJson zero-day (CVE-2026-16723) is being exploited with no patch likely - check for FastJson 1.x and enable SafeMode; EY discloses a breach of a third-party support-ticket platform holding client tax data (extortion claims unverified) - audit your help-desk attachments; and the US and China plan first formal AI talks for September. Plus: Big Tech earnings held over to next week's episode. Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: Axios & TechCrunch (Opus 5), TechTimes & Interconnects (Kimi K3 open weights), Google blog & Axios (Gemini), BleepingComputer & Imperva (FastJson CVE-2026-16723), BleepingComputer & TechRadar Pro (EY), Reuters via CNBC (US-China AI talks).

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.

Right now, your company's executives are probably sitting in some boardroom and they're just aggressively trying to figure out how to integrate this new AI model that can literally reason like a human being. Right. Meanwhile, like three floors down, your IT department is just quietly panicking because the software that's holding your entire payroll data together is this abandoned, unpatchable piece of code from over a decade ago. Yeah, it really is the defining contradiction of modern enterprise tech. I mean, we are building these

towering futuristic skyscrapers right on top of a foundation that is basically crumbling into dust. Exactly. And that's split screen reality. That's our mission for you today on this deep dive. We're drawing from a really dense, fact -packed August 2026 update from Tech Talks with Kinsoft. And we're going to navigate the absolute whiplash between these incredible democratized AI breakthroughs, which are changing the world by the hour, and the, frankly, terrifying forgotten vulnerabilities lurking in the everyday plumbing of our digital infrastructure. Because if you're

only paying attention to the shiny new models, you are entirely missing the hidden catastrophic costs of running them on top of legacy systems. Okay, let's unpack this by starting at the absolute bleeding edge, right? The frontier. Because the future is not just arriving faster than anyone predicted, it's arriving far, far cheaper. It really is. So on July 24th, Anthropic released Claude Opus 5. According to the Kinsoft update, this model claims a new state -of -the -art on frontier reasoning benchmarks. Plus, their behavioral

audits rate it as the most aligned model they've ever shipped. Which is huge on its own. Right. But here is the detail that made me do a literal double -take. The pricing. It is $5 for input and $25 for output per million tokens. What's fascinating here is that this is the exact same price as the previous generation model it replaces. Wow. Yeah, they also released a fast premium mode. But that flat baseline pricing for state of the art reasoning, it's just a massive market disruptor. So before we get into the market impact,

let's ground those numbers for the listener. Because, you know, words like tokens and input versus output, it can sound like abstract math. What does a million tokens actually look like? Okay, so think of a token as roughly three quarters of a word. Okay. So a million tokens is about 750 ,000 words. If you were to feed the AI the entire text of the Harry Potter series, like from the first book to the absolute last, that's roughly a million tokens. Okay, so for five bucks, I can upload the equivalent of seven thick novels for the AI to read and process. Exactly. But

why is the output, like the answer, it actually gives me price five times higher at $25. It really just comes down to the mechanics of how these models come together. You know, input is relatively cheap because the model processes the test you upload in parallel. It just encodes it into its memory space. Right. But output is generative. The model has to predict every single word one by one sequentially. So it takes significantly more computational power and time to generate new text than it does to just read existing text. That makes perfect sense. But the fact that Anthropic

is keeping this price flat. while vastly increasing the model's reasoning capabilities. I mean, that feels like a watershed moment. Oh, absolutely. Like if I priced an AI project six months ago and shelved it because it was just too expensive, is it like I decided not to buy a sports car? And today I woke up to find they're selling them for the price of a bicycle. That's a great way to put it. Yeah. It completely upends the build versus buy math. Right. Because earlier this year, businesses were deciding they couldn't afford to buy API access to frontier models at scale. So they were looking into building smaller,

cheaper workarounds. Makes sense. But today. Capability is spiking exponentially while the cost floor is just dropping out. You quite literally have to tear up your IT budget from six months ago and price it all over again this week. And Anthropic isn't the only one dropping bombs on the market right now. Over that same weekend, China's Moonshot AI released the open weights for their new model, Kimi K3. Yes. The scale here is staggering. 2 .8 trillion parameters. It's the largest open weight release in history.

And we should probably clarify what open weights means mechanically, right? Yeah, let's do that. So unlike Anthropic or OpenAI, where you just rent access to a model through an API, Moonshot is essentially giving developers the core mathematical architecture you know, the weights of Kimi K3. This means an organization can take that massive 2 .8 trillion parameter model and self -host it on their own internal servers. And the immediate assumption is that this is perfect for data sovereignty, right? Like if you are a massive bank or a healthcare

provider with really strict data residency laws, you legally can't send your proprietary data out over the internet to a third -party cloud. So self -hosting Kimi K3 gives you frontier class AI without the data ever leaving your building. But I have to push back on this a bit. Go ahead. We're talking about a model built by China's Moonshot AI. Is a highly regulated U .S. health care provider really going to feel secure downloading and self -hosting a massive Chinese open weight model? Doesn't that just trade like a data privacy risk for a geopolitical risk vector? It absolutely

shifts the risk profile. And, you know, any compliance officer would have a complete field day with it. Right. Even putting geopolitics entirely aside, the infrastructure reality of hosting a 2 .8 trillion parameter model is just a monumental headache. How so? Well, parameters are essentially the artificial synapses of the model. 2 .8 trillion of them requires an unbelievable amount of VRAM, that's video RAM, just to hold the model in memory, let alone actually run inferences on it. So this isn't something I can just, you know, download

onto a high -end corporate server rack in the closet. Not at all. I mean, you were talking about lashing together massive clusters of specialized GPUs. You have to upgrade your data center's cooling infrastructure. Wow. Yeah. The power grid draw alone is staggering. Self -hosting a model this size is no longer just a software project. It is a massive capital expenditure. So the capability might be free, but the infrastructure toll is incredibly heavy. Which perfectly frames the incredibly weird week over at Google. Because keeping up this breakneck pace, whether it's

anthropic optimization, Right. And Google seems to be experiencing a bit of a split screen reality of its own. On July 21st, Google shipped Gemini 3 .6 Flash. Right. And Flash is their designated tier for fast, cheap, workhorse AI. Exactly. And the specs are impressive. It boasts a one million token context window. And they even launched a variant specifically trained for cybersecurity

workloads. So it looks like they are moving at light speed on the surface. But here's where it gets really interesting. Two days later, on July 23rd, Axios drops a report revealing that Google's actual flagship model, Gemini 3 .5 Pro, keeps slipping. It's delayed. And the reasons cited aren't technical bottlenecks. They are organizational friction, morale problems and senior departures at deep. mind. That contrast is so vital for anyone who's mapping out an AI vendor strategy right now. Well, let's dig into that contrast because I look at this and I just see a classic smoke screen. Like, is it possible

that the shipping velocity of these cheap workhorse models like Gemini 3 .6 Flash is really just covering up the fact that Google is stalling out at the actual frontier of research? This raises an important question about how AI development actually works in practice. You know, you have to separate engineering from science. Okay. What do you mean? Well, shipping a flash model, which is taking an existing architecture and making it run faster, cheaper, and with a larger context window that is an engineering optimization problem. It's incredibly difficult, sure, but it's a known

path. So you're basically just refining the engine you already built. Precisely. But pushing the actual frontier of reasoning, like building the flagship Gemini 3 .5 Pro, that requires genuine scientific breakthroughs. It requires researchers discovering entirely new ways to train models. And scientific breakthroughs require a cohesive, highly motivated team working perfectly in sync. Which is hard to do if morale is tanking. Exactly. If morale drops, or... If there is an internal culture war between shipping products fast versus

ensuring model safety, the senior researchers leave. And when the researchers leave, the science just stalls. So if you're a business leader. and you're betting your company's future on a multi -year AI contract, you really cannot just look at the efficiency of the bottom tier. You have to look at the health of the research division. Exactly. The high shipping velocity of those optimization models doesn't guarantee that the actual engine of innovation is healthy. It just highlights how complex and frankly fragile the human element behind these massive tech stacks

really is. Fragile is the perfect word to pivot to the second half of this deep dive. Yeah. Because while everyone is just staring up at the AI frontier, totally mesmerized by what Claude or Gemini can do, the most existential threats to our businesses are actually hiding the boring forgotten plumbing in the Internet. Let's talk about FastJSON. Yes. This is the zero -day vulnerability from the Kinsoft report that should genuinely keep every IT manager awake tonight. The report details a critical, unauthenticated remote code execution flaw in FastJSON, which is Alibaba's widely used

Java JSON library. Right. And it's currently being actively exploited across the financial, healthcare, and retail sectors. So let's break down the mechanics here, because unauthenticated remote code execution, or RCE, it sounds bad, but how does it actually work? So to understand the flaw, you have to understand the software itself. A JSON library is used to parse data. In modern enterprise architecture, you have hundreds of microservices, right? Okay. Basically, servers passing data back and forth to each other constantly.

JSON is the format that data is packaged in. And because Java has been the backbone of enterprise software for 20 years, fast JSON is baked into the absolute deepest plumbing of millions of applications. So it's essentially the mailroom of the server, opening packages of data and just reading what's inside. That's a great way to look at it. Now, an unauthenticated RCE flaw means a hacker doesn't even need to steal a password to get into the system. Wait, really? Really. They simply craft a malicious package of data, a specific string of code, and they just send

it to the server. Wow. Yeah. And when the FastJSON library opens that package, the flaw in the parsing logic tricks the server into immediately executing that malicious code. So the attacker gains complete control from the outside. So what does this all mean? It sounds like... realizing the foundational beams of your company's headquarters are made of rotting wood. But the contractor who built it went out of business a decade ago. That is exactly what it's like. Because if we connect this to the bigger picture, the truly terrifying caveat from the report is that FastJSON 1 .x

is no longer maintained. Right. It's reached its end of life. Yes. Which means a conventional software patch to fix this vulnerability may literally never arrive. So a company running this software can't just, you know. run a quick update and be safe no they can't and it shows that quietly end of life software isn't just a bug it is a distinct vulnerability class of its own when a library reaches end of life the original developers have moved on the cavalry is not coming so what are the options Well, the Kinsoft report notes that the only mitigations

are turning on something called safe mode, which restricts some of the library's functionality to block the exploit, or entirely migrating your systems to a maintained alternative. And migrating away from deeply embedded plumbing like that? I mean, that must be an absolute nightmare. You risk breaking countless dependencies in the code, right? It is incredibly painful and very expensive. It requires thousands of developer hours to rip out a foundational library and replace it. Right. So companies often just leave it. They cross

their fingers and just hope they don't get hit. But right now, attackers are actively exploiting this in the wild. So for you listening to this right now, think about your own organization's tech stack. You need to ask your IT team today, are we running FastJSON 1 .x anywhere in our architecture? And if so, is safe mode turned on? Absolutely. Because if you don't know the answer, the threat actors currently hunting through the financial sector, they will eventually find out for you. And we shouldn't just focus on forgotten, unmaintained code either. The everyday, perfectly

functioning systems that we completely take for granted, those are increasingly becoming the weakest links to. This is where we get into human behavior. Yeah. And the Ernst & Young story from the report gave me actual chills because I instantly recognized my own bad habits here. Wow, we all do it. Right. So on July 15th... EY notified U .S. regulators that an unauthorized party spent a fortnight from late March to mid -April inside a third -party IT support ticket platform. Just poking around. Exactly. And while they were inside

this help desk system, the hackers downloaded client tax and financial data. Now, an unverified extortion group is claiming responsibility and threatening to leak the files. And the critical detail here is why highly sensitive client tax data was sitting inside an IT support ticket platform in the first place. Right, because this is just human nature, isn't it? Think about the last time your accounting software glitched. You're trying to meet a deadline, a crucial tax file won't open, or the formatting just breaks. You're stressed. You're in a hurry. Exactly. You don't think about the incredibly sensitive

data inside the file. You take a screenshot or you attach the broken file itself to a Zendesk or an IT support ticket and you write, hey, fix this. We are basically using a glass wall as a bank vault just because it's convenient. bypassed millions of dollars of corporate cybersecurity to drop a client's social security number into a glorified inbox. You have perfectly described how shadow data repositories are born. Shadow data repositories. Every major organization has its official, highly guarded vaults, the CRM,

the main financial database. The systems get audited annually. They have strict access controls, encryption, the works. But a third -party IT help desk platform, that's just... Considered operational overhead. It rarely, if ever, falls under the scope of rigorous data protection reviews. Because the security team just assumes it's an inbox for, you know, my printer's jammed complaints. Exactly. But over years of operation, because of that very human nature you described, that help desk system accumulates tens of thousands

of attachments. Unencrypted tax files, payroll spreadsheets, internal strategy documents. passwords sent in plain text. Wow. Yeah, the support ticket platform silently becomes a shadow database that actually contains more raw, sensitive information than the heavily fortified CRM. And worse, as we saw with EY, these systems are increasingly outsourced to third -party vendors, which just multiplies the risk factor. It makes you wonder, what is sitting in your company's help desk attachments right now? Yeah. Are your outsourced IT support

agents inadvertently holding the keys to the kingdom? Almost certainly. And when you step back and look at the whole board we've discussed today, I mean, the unbridled pace of unregulated AI scaling up, the unpatchable legacy code woven into our financial infrastructure, and these massive shadow databases hiding in plain sight the sheer chaos of the tech landscape has reached a boiling point. The private sector is clearly struggling to manage this risk independently. Which means we are at the exact moment where governments are being forced to step in. They're

pulling this entire conversation out of the server room and onto the global stage. It's a huge shift. Reuters is reporting that Washington and Beijing are going to hold their first dedicated AI talks in September. This is slated just ahead of Xi Jinping's planned US visit. Right. And the agenda is heavy. They're going to cover monumental risks like the military use of AI and potential AI -driven attacks on critical infrastructure. And we need to look at this strictly through the lens of market impact and compliance because

the mere existence of these formal bilateral channels between Washington and Beijing, it signals a fundamental shift in how artificial intelligence is going to be treated globally. We are moving from tech bros debating ethics on conference panels to actual high -stakes statecraft. Are we about to see AI treated like nuclear material on the world stage? It's certainly heading in a highly regulated direction. And where statecraft goes, massive business compliance inevitably follows. Right. The establishment of these talks tells the market that incoming export rules are

going to get significantly tighter. We are going to see severe model access restrictions. Governments will begin dictating who is legally allowed. to buy rent or self -host those massive frontier models we talked about the beginning of the hour so the red tape is coming big time it signals that massive incredibly expensive compliance obligations are coming for any business utilizing this technology at scale The economics of overhauling an enterprise tech stack while simultaneously funding frontier AI research and then paying

for all this new government compliance. It's staggering. It is staggering. And it's actually the next economic shoe to drop in this story. Oh, with the earnings reports? Yes. Next week brings the big tech June quarter earnings. And the market is going to be intensely focused on the AI capital expenditures from Microsoft, Meta, Apple and Amazon. The sheer amount of money they are pouring into this infrastructure. historically unprecedented. So those earnings reports will tell us if the economic reality and the revenue generation are actually keeping pace with the

geopolitical and technical chaos we are seeing right now. It is a phenomenal amount of information to process and it really paints a picture of an industry moving way faster than its own guardrails. Absolutely. So as we wrap up this deep dive, let's just take a breath and look at the whiplash journey we just went on. Let's do it. In a single update, we've gone from the awe -inspiring reality of $5 AI models that can reason like humans. all the way down into the terrifying, rotting floorboards of unpatchable Java flaws and hijacked

IT support tickets. The contrast is staggering. But I want to leave you with a final thought to mull over, one that connects all these seemingly disparate threads we've discussed today. Okay, let's hear it. We know that AI models like Opus 5 are getting exponentially cheaper and wildly better at frontier reasoning. And we now know that our enterprise systems are filled with unpatchable plumbing like FastJSON and massive shadow repositories. of tax documents sitting in forgotten help desk attachments. Oh, they can see the collision you're setting up here. What happens when these two

forces inevitably collide? Up until now, we've relied on the fact that hackers are human. They only have so much time to manually sift through thousands of help desk attachments or manually probe legacy systems for fast JSON vulnerabilities. But what happens when cheap reasoning AI is unleashed by bad actors? Will we soon see these frontier models used to automatically hunt down those forgotten help desk attachments and exploit those unpatchable libraries at a massive automated scale? You essentially have an automated tireless

hunter probing every inch of that crumbling foundation we talked about. Well, that is a genuinely unsettling thought to end on, but it is exactly why we need to understand the whole picture. not just the shiny new releases. The landscape is moving incredibly fast, and the foundations are way more fragile than they appear. Thank you for joining us on this deep dive. Stay curious, go check your own tech stack seriously, ask about that fast JSON library, and check your IT ticket attachments today, and join us next time as we keep unpacking the future.

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