Latest / Tech Talks With Kinsoft / Last Week in Tech – Claude Opus 5, a Patchless Java Zero-Day, and EY's Support-Desk Breach

Transcript

Last Week in Tech – Claude Opus 5, a Patchless Java Zero-Day, and EY's Support-Desk Breach · Tech Talks With Kinsoft

• 0:00
  Right now, your company's executives are probably
• 0:03
  sitting in some boardroom and they're just aggressively
• 0:07
  trying to figure out how to integrate this new
• 0:08
  AI model that can literally reason like a human
• 0:12
  being. Right. Meanwhile, like three floors down,
• 0:15
  your IT department is just quietly panicking
• 0:18
  because the software that's holding your entire
• 0:20
  payroll data together is this abandoned, unpatchable
• 0:23
  piece of code from over a decade ago. Yeah, it
• 0:26
  really is the defining contradiction of modern
• 0:29
  enterprise tech. I mean, we are building these
• 0:31
  towering futuristic skyscrapers right on top
• 0:35
  of a foundation that is basically crumbling into
• 0:37
  dust. Exactly. And that's split screen reality.
• 0:40
  That's our mission for you today on this deep
• 0:42
  dive. We're drawing from a really dense, fact
• 0:45
  -packed August 2026 update from Tech Talks with
• 0:48
  Kinsoft. And we're going to navigate the absolute
• 0:50
  whiplash between these incredible democratized
• 0:53
  AI breakthroughs, which are changing the world
• 0:55
  by the hour, and the, frankly, terrifying forgotten
• 0:58
  vulnerabilities lurking in the everyday plumbing
• 1:01
  of our digital infrastructure. Because if you're
• 1:03
  only paying attention to the shiny new models,
• 1:05
  you are entirely missing the hidden catastrophic
• 1:08
  costs of running them on top of legacy systems.
• 1:11
  Okay, let's unpack this by starting at the absolute
• 1:14
  bleeding edge, right? The frontier. Because the
• 1:16
  future is not just arriving faster than anyone
• 1:19
  predicted, it's arriving far, far cheaper. It
• 1:21
  really is. So on July 24th, Anthropic released
• 1:24
  Claude Opus 5. According to the Kinsoft update,
• 1:27
  this model claims a new state -of -the -art on
• 1:30
  frontier reasoning benchmarks. Plus, their behavioral
• 1:33
  audits rate it as the most aligned model they've
• 1:36
  ever shipped. Which is huge on its own. Right.
• 1:38
  But here is the detail that made me do a literal
• 1:41
  double -take. The pricing. It is $5 for input
• 1:44
  and $25 for output per million tokens. What's
• 1:48
  fascinating here is that this is the exact same
• 1:50
  price as the previous generation model it replaces.
• 1:53
  Wow. Yeah, they also released a fast premium
• 1:55
  mode. But that flat baseline pricing for state
• 1:58
  of the art reasoning, it's just a massive market
• 2:01
  disruptor. So before we get into the market impact,
• 2:04
  let's ground those numbers for the listener.
• 2:06
  Because, you know, words like tokens and input
• 2:09
  versus output, it can sound like abstract math.
• 2:11
  What does a million tokens actually look like?
• 2:14
  Okay, so think of a token as roughly three quarters
• 2:16
  of a word. Okay. So a million tokens is about
• 2:19
  750 ,000 words. If you were to feed the AI the
• 2:23
  entire text of the Harry Potter series, like
• 2:26
  from the first book to the absolute last, that's
• 2:28
  roughly a million tokens. Okay, so for five bucks,
• 2:31
  I can upload the equivalent of seven thick novels
• 2:33
  for the AI to read and process. Exactly. But
• 2:36
  why is the output, like the answer, it actually
• 2:38
  gives me price five times higher at $25. It really
• 2:41
  just comes down to the mechanics of how these
• 2:43
  models come together. You know, input is relatively
• 2:45
  cheap because the model processes the test you
• 2:48
  upload in parallel. It just encodes it into its
• 2:51
  memory space. Right. But output is generative.
• 2:53
  The model has to predict every single word one
• 2:56
  by one sequentially. So it takes significantly
• 2:59
  more computational power and time to generate
• 3:01
  new text than it does to just read existing text.
• 3:04
  That makes perfect sense. But the fact that Anthropic
• 3:08
  is keeping this price flat. while vastly increasing
• 3:11
  the model's reasoning capabilities. I mean, that
• 3:14
  feels like a watershed moment. Oh, absolutely.
• 3:17
  Like if I priced an AI project six months ago
• 3:19
  and shelved it because it was just too expensive,
• 3:21
  is it like I decided not to buy a sports car?
• 3:24
  And today I woke up to find they're selling them
• 3:26
  for the price of a bicycle. That's a great way
• 3:27
  to put it. Yeah. It completely upends the build
• 3:30
  versus buy math. Right. Because earlier this
• 3:33
  year, businesses were deciding they couldn't
• 3:35
  afford to buy API access to frontier models at
• 3:37
  scale. So they were looking into building smaller,
• 3:39
  cheaper workarounds. Makes sense. But today.
• 3:43
  Capability is spiking exponentially while the
• 3:45
  cost floor is just dropping out. You quite literally
• 3:47
  have to tear up your IT budget from six months
• 3:49
  ago and price it all over again this week. And
• 3:51
  Anthropic isn't the only one dropping bombs on
• 3:53
  the market right now. Over that same weekend,
• 3:56
  China's Moonshot AI released the open weights
• 3:59
  for their new model, Kimi K3. Yes. The scale
• 4:03
  here is staggering. 2 .8 trillion parameters.
• 4:07
  It's the largest open weight release in history.
• 4:10
  And we should probably clarify what open weights
• 4:12
  means mechanically, right? Yeah, let's do that.
• 4:14
  So unlike Anthropic or OpenAI, where you just
• 4:17
  rent access to a model through an API, Moonshot
• 4:20
  is essentially giving developers the core mathematical
• 4:23
  architecture you know, the weights of Kimi K3.
• 4:27
  This means an organization can take that massive
• 4:30
  2 .8 trillion parameter model and self -host
• 4:33
  it on their own internal servers. And the immediate
• 4:36
  assumption is that this is perfect for data sovereignty,
• 4:39
  right? Like if you are a massive bank or a healthcare
• 4:42
  provider with really strict data residency laws,
• 4:45
  you legally can't send your proprietary data
• 4:49
  out over the internet to a third -party cloud.
• 4:51
  So self -hosting Kimi K3 gives you frontier class
• 4:54
  AI without the data ever leaving your building.
• 4:57
  But I have to push back on this a bit. Go ahead.
• 4:59
  We're talking about a model built by China's
• 5:01
  Moonshot AI. Is a highly regulated U .S. health
• 5:04
  care provider really going to feel secure downloading
• 5:06
  and self -hosting a massive Chinese open weight
• 5:08
  model? Doesn't that just trade like a data privacy
• 5:11
  risk for a geopolitical risk vector? It absolutely
• 5:14
  shifts the risk profile. And, you know, any compliance
• 5:17
  officer would have a complete field day with
• 5:19
  it. Right. Even putting geopolitics entirely
• 5:22
  aside, the infrastructure reality of hosting
• 5:25
  a 2 .8 trillion parameter model is just a monumental
• 5:28
  headache. How so? Well, parameters are essentially
• 5:31
  the artificial synapses of the model. 2 .8 trillion
• 5:35
  of them requires an unbelievable amount of VRAM,
• 5:38
  that's video RAM, just to hold the model in memory,
• 5:41
  let alone actually run inferences on it. So this
• 5:44
  isn't something I can just, you know, download
• 5:45
  onto a high -end corporate server rack in the
• 5:47
  closet. Not at all. I mean, you were talking
• 5:50
  about lashing together massive clusters of specialized
• 5:53
  GPUs. You have to upgrade your data center's
• 5:55
  cooling infrastructure. Wow. Yeah. The power
• 5:58
  grid draw alone is staggering. Self -hosting
• 6:00
  a model this size is no longer just a software
• 6:02
  project. It is a massive capital expenditure.
• 6:05
  So the capability might be free, but the infrastructure
• 6:08
  toll is incredibly heavy. Which perfectly frames
• 6:11
  the incredibly weird week over at Google. Because
• 6:14
  keeping up this breakneck pace, whether it's
• 6:16
  anthropic optimization, Right. And Google seems
• 6:25
  to be experiencing a bit of a split screen reality
• 6:27
  of its own. On July 21st, Google shipped Gemini
• 6:31
  3 .6 Flash. Right. And Flash is their designated
• 6:35
  tier for fast, cheap, workhorse AI. Exactly.
• 6:39
  And the specs are impressive. It boasts a one
• 6:41
  million token context window. And they even launched
• 6:44
  a variant specifically trained for cybersecurity
• 6:47
  workloads. So it looks like they are moving at
• 6:50
  light speed on the surface. But here's where
• 6:52
  it gets really interesting. Two days later, on
• 6:54
  July 23rd, Axios drops a report revealing that
• 6:56
  Google's actual flagship model, Gemini 3 .5 Pro,
• 7:00
  keeps slipping. It's delayed. And the reasons
• 7:02
  cited aren't technical bottlenecks. They are
• 7:04
  organizational friction, morale problems and
• 7:07
  senior departures at deep. mind. That contrast
• 7:09
  is so vital for anyone who's mapping out an AI
• 7:11
  vendor strategy right now. Well, let's dig into
• 7:14
  that contrast because I look at this and I just
• 7:16
  see a classic smoke screen. Like, is it possible
• 7:20
  that the shipping velocity of these cheap workhorse
• 7:22
  models like Gemini 3 .6 Flash is really just
• 7:25
  covering up the fact that Google is stalling
• 7:27
  out at the actual frontier of research? This
• 7:30
  raises an important question about how AI development
• 7:32
  actually works in practice. You know, you have
• 7:35
  to separate engineering from science. Okay. What
• 7:38
  do you mean? Well, shipping a flash model, which
• 7:40
  is taking an existing architecture and making
• 7:42
  it run faster, cheaper, and with a larger context
• 7:46
  window that is an engineering optimization problem.
• 7:49
  It's incredibly difficult, sure, but it's a known
• 7:52
  path. So you're basically just refining the engine
• 7:54
  you already built. Precisely. But pushing the
• 7:57
  actual frontier of reasoning, like building the
• 7:59
  flagship Gemini 3 .5 Pro, that requires genuine
• 8:02
  scientific breakthroughs. It requires researchers
• 8:05
  discovering entirely new ways to train models.
• 8:08
  And scientific breakthroughs require a cohesive,
• 8:11
  highly motivated team working perfectly in sync.
• 8:13
  Which is hard to do if morale is tanking. Exactly.
• 8:16
  If morale drops, or... If there is an internal
• 8:19
  culture war between shipping products fast versus
• 8:22
  ensuring model safety, the senior researchers
• 8:25
  leave. And when the researchers leave, the science
• 8:27
  just stalls. So if you're a business leader.
• 8:30
  and you're betting your company's future on a
• 8:32
  multi -year AI contract, you really cannot just
• 8:36
  look at the efficiency of the bottom tier. You
• 8:38
  have to look at the health of the research division.
• 8:41
  Exactly. The high shipping velocity of those
• 8:43
  optimization models doesn't guarantee that the
• 8:46
  actual engine of innovation is healthy. It just
• 8:49
  highlights how complex and frankly fragile the
• 8:52
  human element behind these massive tech stacks
• 8:54
  really is. Fragile is the perfect word to pivot
• 8:57
  to the second half of this deep dive. Yeah. Because
• 8:59
  while everyone is just staring up at the AI frontier,
• 9:01
  totally mesmerized by what Claude or Gemini can
• 9:04
  do, the most existential threats to our businesses
• 9:06
  are actually hiding the boring forgotten plumbing
• 9:09
  in the Internet. Let's talk about FastJSON. Yes.
• 9:12
  This is the zero -day vulnerability from the
• 9:14
  Kinsoft report that should genuinely keep every
• 9:17
  IT manager awake tonight. The report details
• 9:19
  a critical, unauthenticated remote code execution
• 9:22
  flaw in FastJSON, which is Alibaba's widely used
• 9:27
  Java JSON library. Right. And it's currently
• 9:30
  being actively exploited across the financial,
• 9:33
  healthcare, and retail sectors. So let's break
• 9:36
  down the mechanics here, because unauthenticated
• 9:38
  remote code execution, or RCE, it sounds bad,
• 9:42
  but how does it actually work? So to understand
• 9:44
  the flaw, you have to understand the software
• 9:46
  itself. A JSON library is used to parse data.
• 9:50
  In modern enterprise architecture, you have hundreds
• 9:52
  of microservices, right? Okay. Basically, servers
• 9:54
  passing data back and forth to each other constantly.
• 9:57
  JSON is the format that data is packaged in.
• 10:00
  And because Java has been the backbone of enterprise
• 10:02
  software for 20 years, fast JSON is baked into
• 10:05
  the absolute deepest plumbing of millions of
• 10:08
  applications. So it's essentially the mailroom
• 10:10
  of the server, opening packages of data and just
• 10:13
  reading what's inside. That's a great way to
• 10:15
  look at it. Now, an unauthenticated RCE flaw
• 10:17
  means a hacker doesn't even need to steal a password
• 10:20
  to get into the system. Wait, really? Really.
• 10:23
  They simply craft a malicious package of data,
• 10:26
  a specific string of code, and they just send
• 10:28
  it to the server. Wow. Yeah. And when the FastJSON
• 10:31
  library opens that package, the flaw in the parsing
• 10:34
  logic tricks the server into immediately executing
• 10:37
  that malicious code. So the attacker gains complete
• 10:40
  control from the outside. So what does this all
• 10:43
  mean? It sounds like... realizing the foundational
• 10:46
  beams of your company's headquarters are made
• 10:48
  of rotting wood. But the contractor who built
• 10:51
  it went out of business a decade ago. That is
• 10:53
  exactly what it's like. Because if we connect
• 10:55
  this to the bigger picture, the truly terrifying
• 10:57
  caveat from the report is that FastJSON 1 .x
• 11:01
  is no longer maintained. Right. It's reached
• 11:03
  its end of life. Yes. Which means a conventional
• 11:05
  software patch to fix this vulnerability may
• 11:07
  literally never arrive. So a company running
• 11:10
  this software can't just, you know. run a quick
• 11:12
  update and be safe no they can't and it shows
• 11:14
  that quietly end of life software isn't just
• 11:17
  a bug it is a distinct vulnerability class of
• 11:20
  its own when a library reaches end of life the
• 11:23
  original developers have moved on the cavalry
• 11:26
  is not coming so what are the options Well, the
• 11:29
  Kinsoft report notes that the only mitigations
• 11:31
  are turning on something called safe mode, which
• 11:33
  restricts some of the library's functionality
• 11:35
  to block the exploit, or entirely migrating your
• 11:39
  systems to a maintained alternative. And migrating
• 11:42
  away from deeply embedded plumbing like that?
• 11:45
  I mean, that must be an absolute nightmare. You
• 11:48
  risk breaking countless dependencies in the code,
• 11:51
  right? It is incredibly painful and very expensive.
• 11:55
  It requires thousands of developer hours to rip
• 11:58
  out a foundational library and replace it. Right.
• 12:01
  So companies often just leave it. They cross
• 12:03
  their fingers and just hope they don't get hit.
• 12:04
  But right now, attackers are actively exploiting
• 12:08
  this in the wild. So for you listening to this
• 12:10
  right now, think about your own organization's
• 12:12
  tech stack. You need to ask your IT team today,
• 12:15
  are we running FastJSON 1 .x anywhere in our
• 12:19
  architecture? And if so, is safe mode turned
• 12:22
  on? Absolutely. Because if you don't know the
• 12:24
  answer, the threat actors currently hunting through
• 12:26
  the financial sector, they will eventually find
• 12:28
  out for you. And we shouldn't just focus on forgotten,
• 12:31
  unmaintained code either. The everyday, perfectly
• 12:33
  functioning systems that we completely take for
• 12:36
  granted, those are increasingly becoming the
• 12:38
  weakest links to. This is where we get into human
• 12:41
  behavior. Yeah. And the Ernst & Young story from
• 12:44
  the report gave me actual chills because I instantly
• 12:46
  recognized my own bad habits here. Wow, we all
• 12:48
  do it. Right. So on July 15th... EY notified
• 12:52
  U .S. regulators that an unauthorized party spent
• 12:55
  a fortnight from late March to mid -April inside
• 12:58
  a third -party IT support ticket platform. Just
• 13:01
  poking around. Exactly. And while they were inside
• 13:04
  this help desk system, the hackers downloaded
• 13:06
  client tax and financial data. Now, an unverified
• 13:10
  extortion group is claiming responsibility and
• 13:13
  threatening to leak the files. And the critical
• 13:15
  detail here is why highly sensitive client tax
• 13:18
  data was sitting inside an IT support ticket
• 13:21
  platform in the first place. Right, because this
• 13:22
  is just human nature, isn't it? Think about the
• 13:25
  last time your accounting software glitched.
• 13:27
  You're trying to meet a deadline, a crucial tax
• 13:29
  file won't open, or the formatting just breaks.
• 13:31
  You're stressed. You're in a hurry. Exactly.
• 13:33
  You don't think about the incredibly sensitive
• 13:35
  data inside the file. You take a screenshot or
• 13:39
  you attach the broken file itself to a Zendesk
• 13:41
  or an IT support ticket and you write, hey, fix
• 13:43
  this. We are basically using a glass wall as
• 13:47
  a bank vault just because it's convenient. bypassed
• 13:50
  millions of dollars of corporate cybersecurity
• 13:52
  to drop a client's social security number into
• 13:55
  a glorified inbox. You have perfectly described
• 13:58
  how shadow data repositories are born. Shadow
• 14:01
  data repositories. Every major organization has
• 14:05
  its official, highly guarded vaults, the CRM,
• 14:08
  the main financial database. The systems get
• 14:10
  audited annually. They have strict access controls,
• 14:13
  encryption, the works. But a third -party IT
• 14:16
  help desk platform, that's just... Considered
• 14:19
  operational overhead. It rarely, if ever, falls
• 14:22
  under the scope of rigorous data protection reviews.
• 14:25
  Because the security team just assumes it's an
• 14:26
  inbox for, you know, my printer's jammed complaints.
• 14:29
  Exactly. But over years of operation, because
• 14:32
  of that very human nature you described, that
• 14:35
  help desk system accumulates tens of thousands
• 14:38
  of attachments. Unencrypted tax files, payroll
• 14:41
  spreadsheets, internal strategy documents. passwords
• 14:45
  sent in plain text. Wow. Yeah, the support ticket
• 14:48
  platform silently becomes a shadow database that
• 14:51
  actually contains more raw, sensitive information
• 14:53
  than the heavily fortified CRM. And worse, as
• 14:57
  we saw with EY, these systems are increasingly
• 14:59
  outsourced to third -party vendors, which just
• 15:02
  multiplies the risk factor. It makes you wonder,
• 15:05
  what is sitting in your company's help desk attachments
• 15:07
  right now? Yeah. Are your outsourced IT support
• 15:10
  agents inadvertently holding the keys to the
• 15:12
  kingdom? Almost certainly. And when you step
• 15:14
  back and look at the whole board we've discussed
• 15:16
  today, I mean, the unbridled pace of unregulated
• 15:19
  AI scaling up, the unpatchable legacy code woven
• 15:23
  into our financial infrastructure, and these
• 15:25
  massive shadow databases hiding in plain sight
• 15:27
  the sheer chaos of the tech landscape has reached
• 15:30
  a boiling point. The private sector is clearly
• 15:32
  struggling to manage this risk independently.
• 15:35
  Which means we are at the exact moment where
• 15:37
  governments are being forced to step in. They're
• 15:40
  pulling this entire conversation out of the server
• 15:43
  room and onto the global stage. It's a huge shift.
• 15:46
  Reuters is reporting that Washington and Beijing
• 15:48
  are going to hold their first dedicated AI talks
• 15:52
  in September. This is slated just ahead of Xi
• 15:55
  Jinping's planned US visit. Right. And the agenda
• 15:59
  is heavy. They're going to cover monumental risks
• 16:01
  like the military use of AI and potential AI
• 16:05
  -driven attacks on critical infrastructure. And
• 16:08
  we need to look at this strictly through the
• 16:09
  lens of market impact and compliance because
• 16:11
  the mere existence of these formal bilateral
• 16:13
  channels between Washington and Beijing, it signals
• 16:17
  a fundamental shift in how artificial intelligence
• 16:19
  is going to be treated globally. We are moving
• 16:22
  from tech bros debating ethics on conference
• 16:24
  panels to actual high -stakes statecraft. Are
• 16:27
  we about to see AI treated like nuclear material
• 16:30
  on the world stage? It's certainly heading in
• 16:32
  a highly regulated direction. And where statecraft
• 16:35
  goes, massive business compliance inevitably
• 16:37
  follows. Right. The establishment of these talks
• 16:40
  tells the market that incoming export rules are
• 16:43
  going to get significantly tighter. We are going
• 16:46
  to see severe model access restrictions. Governments
• 16:50
  will begin dictating who is legally allowed.
• 16:53
  to buy rent or self -host those massive frontier
• 16:56
  models we talked about the beginning of the hour
• 16:58
  so the red tape is coming big time it signals
• 17:01
  that massive incredibly expensive compliance
• 17:03
  obligations are coming for any business utilizing
• 17:06
  this technology at scale The economics of overhauling
• 17:10
  an enterprise tech stack while simultaneously
• 17:12
  funding frontier AI research and then paying
• 17:15
  for all this new government compliance. It's
• 17:17
  staggering. It is staggering. And it's actually
• 17:20
  the next economic shoe to drop in this story.
• 17:22
  Oh, with the earnings reports? Yes. Next week
• 17:24
  brings the big tech June quarter earnings. And
• 17:27
  the market is going to be intensely focused on
• 17:29
  the AI capital expenditures from Microsoft, Meta,
• 17:32
  Apple and Amazon. The sheer amount of money they
• 17:36
  are pouring into this infrastructure. historically
• 17:38
  unprecedented. So those earnings reports will
• 17:41
  tell us if the economic reality and the revenue
• 17:44
  generation are actually keeping pace with the
• 17:46
  geopolitical and technical chaos we are seeing
• 17:49
  right now. It is a phenomenal amount of information
• 17:51
  to process and it really paints a picture of
• 17:54
  an industry moving way faster than its own guardrails.
• 17:58
  Absolutely. So as we wrap up this deep dive,
• 18:01
  let's just take a breath and look at the whiplash
• 18:03
  journey we just went on. Let's do it. In a single
• 18:05
  update, we've gone from the awe -inspiring reality
• 18:08
  of $5 AI models that can reason like humans.
• 18:12
  all the way down into the terrifying, rotting
• 18:15
  floorboards of unpatchable Java flaws and hijacked
• 18:18
  IT support tickets. The contrast is staggering.
• 18:21
  But I want to leave you with a final thought
• 18:22
  to mull over, one that connects all these seemingly
• 18:25
  disparate threads we've discussed today. Okay,
• 18:27
  let's hear it. We know that AI models like Opus
• 18:30
  5 are getting exponentially cheaper and wildly
• 18:32
  better at frontier reasoning. And we now know
• 18:34
  that our enterprise systems are filled with unpatchable
• 18:38
  plumbing like FastJSON and massive shadow repositories.
• 18:41
  of tax documents sitting in forgotten help desk
• 18:44
  attachments. Oh, they can see the collision you're
• 18:47
  setting up here. What happens when these two
• 18:49
  forces inevitably collide? Up until now, we've
• 18:52
  relied on the fact that hackers are human. They
• 18:54
  only have so much time to manually sift through
• 18:56
  thousands of help desk attachments or manually
• 18:59
  probe legacy systems for fast JSON vulnerabilities.
• 19:03
  But what happens when cheap reasoning AI is unleashed
• 19:07
  by bad actors? Will we soon see these frontier
• 19:10
  models used to automatically hunt down those
• 19:12
  forgotten help desk attachments and exploit those
• 19:15
  unpatchable libraries at a massive automated
• 19:18
  scale? You essentially have an automated tireless
• 19:26
  hunter probing every inch of that crumbling foundation
• 19:28
  we talked about. Well, that is a genuinely unsettling
• 19:31
  thought to end on, but it is exactly why we need
• 19:34
  to understand the whole picture. not just the
• 19:37
  shiny new releases. The landscape is moving incredibly
• 19:39
  fast, and the foundations are way more fragile
• 19:42
  than they appear. Thank you for joining us on
• 19:44
  this deep dive. Stay curious, go check your own
• 19:46
  tech stack seriously, ask about that fast JSON
• 19:49
  library, and check your IT ticket attachments
• 19:51
  today, and join us next time as we keep unpacking
• 19:54
  the future.