Latest / Tech Talks With Kinsoft / Last Week in Tech - Nvidia Buys Hugging Face, GPT-6 Crosses the Cyber Line, and Seven Exploited Flaws in a Day

Transcript

Last Week in Tech - Nvidia Buys Hugging Face, GPT-6 Crosses the Cyber Line, and Seven Exploited Flaws in a Day · Tech Talks With Kinsoft

• 0:00
  Have you ever been sitting at your desk and you
• 0:02
  just feel this, I don't know, this profound sense
• 0:05
  of whiplash? Oh, absolutely. All the time. Right.
• 0:08
  Like you're reading an article in one browser
• 0:10
  tab about some trillion dollar leap in artificial
• 0:13
  intelligence. Yeah. You know, a new model that
• 0:15
  feels like literal science fiction. Yeah. Something
• 0:18
  that's just blowing every benchmark out of the
• 0:20
  water. Exactly. And then you click over to another
• 0:22
  tab and there is a notification telling you that
• 0:25
  you forgot to update your basic web browser.
• 0:28
  It's just so frustrating. It is. It's this bizarre
• 0:31
  cognitive dissonance where, you know, we are
• 0:34
  living in the future, but we are still chipping
• 0:36
  over the digital extension cords of the present.
• 0:38
  That is well, I mean, it's really the defining
• 0:40
  tension of our current technological era. We
• 0:43
  are building these towering, incredibly sophisticated
• 0:46
  skyscrapers of machine intelligence. But we are
• 0:50
  essentially resting them on a foundation of.
• 0:56
  I was feeling this exact whiplash yesterday when
• 0:59
  I was reading through a tech dispatch from early
• 1:01
  September 2026. The whole document captures this
• 1:04
  tension so perfectly. I knew we had to make it
• 1:08
  the focus of today's depth dive. That's a great
• 1:10
  source to dig into. Yeah, we are looking at a
• 1:12
  single week in tech that was split like cleanly
• 1:15
  down the middle. In the first half, we have the
• 1:17
  massive money and models fueling the AI boom.
• 1:20
  And in the second half, we have the immediate.
• 1:23
  real -world cyber vulnerabilities that are falling
• 1:26
  out of it landing right on your desk. Okay, let's
• 1:29
  unpack this. Because trying to navigate a landscape
• 1:32
  moving this fast feels, honestly, almost impossible.
• 1:35
  It really does. But I think the trick here is
• 1:38
  to follow the money first. Okay, follow the money.
• 1:40
  Yeah, because to understand the daily security
• 1:42
  headaches you and your IT teams are dealing with,
• 1:44
  you really have to understand the macro. We need
• 1:47
  to trace the billions of dollars currently building
• 1:49
  the physical infrastructure of AI. Then we can
• 1:52
  look at the terrifying new capabilities of the
• 1:55
  frontier models that this money is actually producing.
• 1:57
  And finally, connect all of that high -level
• 2:01
  architecture down to the vulnerabilities actively
• 2:04
  being exploited today. It's a single continuous
• 2:08
  loop. I love that approach. So before we even
• 2:11
  look at what AI is doing to cybersecurity, we
• 2:13
  really have to wrap our heads around the sheer
• 2:15
  scale of the infrastructure being built. Because
• 2:17
  the numbers are huge. The numbers in the source
• 2:20
  are just staggering. Let's start with NVIDIA.
• 2:22
  They are acquiring Hugging Face for $12 .3 billion.
• 2:28
  That is a massive acquisition. It is. And if
• 2:31
  you aren't like a machine learning engineer,
• 2:33
  that might just sound like another tech acquisition.
• 2:35
  But Hugging Face is essentially the public square
• 2:37
  of open AI models. Yeah, it's completely foundational
• 2:40
  to the community. Right. It's a platform hosting
• 2:42
  18 million developers and over 3 million models.
• 2:46
  And NVIDIA's CEO, Jensen Huang, made a very public
• 2:50
  point of... promising that Hugging Face will
• 2:52
  remain completely neutral, you know, multi -cloud
• 2:55
  and open -weight. Which is a big promise. It
• 2:57
  is. And if I understand open -weight correctly,
• 3:00
  that essentially means the core architecture
• 3:02
  of these models is out there for developers to
• 3:05
  download, run, and, you know, tinker with on
• 3:09
  their own hardware rather than being locked behind
• 3:11
  some corporate API. That is the crucial distinction,
• 3:14
  yeah. Open weight models are really the democratic
• 3:17
  counterweight to the massive closed models run
• 3:22
  by companies like OpenAI or Google. Right, because
• 3:24
  you can actually see what's under the hood. Exactly.
• 3:27
  And NVIDIA is fully aware that Hugging Face's
• 3:29
  entire value proposition relies on this neutrality.
• 3:33
  If developers feel like the platform is just
• 3:35
  becoming, I don't know, a funnel to force them
• 3:38
  into buying NVIDIA hardware, they will just migrate
• 3:40
  elsewhere. I hear that, but I have to push back
• 3:43
  a bit on the idea that things won't change. Oh,
• 3:45
  I'm sure they will, in some ways. Right, because
• 3:47
  if Hugging Face is the neutral Switzerland of
• 3:49
  the open -source AI world, NVIDIA just bought
• 3:52
  Switzerland for almost $13 billion. Do we really
• 3:55
  expect them not to plant their flag on the banks?
• 3:58
  Well, that's the billion -dollar question, isn't
• 4:00
  it? Literally. I mean, 18 months from now, when
• 4:02
  the regulators finally sign off, the licensing
• 4:04
  terms could easily shift. And the source makes
• 4:07
  a fantastic practical point here for anyone running
• 4:09
  a business. If your company relies on an AI model,
• 4:12
  you need to write down exactly where it came
• 4:14
  from today. Yeah, it is the shadow IT problem,
• 4:17
  but supersized for the AI era. Oh, that's a good
• 4:21
  way to put it. Right, because 10 years ago, the
• 4:23
  problem was employees spinning up rogue Dropbox
• 4:27
  accounts. Today, it is a contractor who built
• 4:30
  your customer service chatbot using some open
• 4:33
  weight model pulled from Hugging Face. And nobody
• 4:35
  knows which one it is. Exactly. Nobody in your
• 4:38
  company actually documented which version it
• 4:41
  is or what license it operates under. If you
• 4:43
  have a dependency you cannot name, you have a
• 4:45
  vulnerability you cannot manage. And it is not
• 4:47
  just software dependencies we have to worry about,
• 4:49
  right? It is the physical hardware constraint.
• 4:51
  The dispatch highlights Broadcom's third quarter
• 4:54
  numbers for financial year 2026. And they are,
• 4:57
  well, they're eye watering. They really are.
• 5:00
  $29 .59 billion in total revenue. But the critical
• 5:04
  line item is their AI semiconductor revenue.
• 5:07
  $16 .7 billion for the quarter. Wow. Yeah. That
• 5:12
  is up 221 % year on year. When you see a 221
• 5:17
  % increase in a hardware sector, you are looking
• 5:20
  at the bottleneck of the entire industry. The
• 5:22
  absolute bottleneck. Because Broadcom makes these
• 5:26
  specialized networking chips and custom silicon
• 5:28
  that allow tens of thousands of GPUs to talk
• 5:32
  to each other in a data center. That $16 .7 billion
• 5:35
  is an abstract wealth. It represents physical
• 5:38
  capacity being built out at a frantic pace. Like
• 5:41
  actual metal and wire. Yes. It represents copper
• 5:44
  wiring, immense power grid consumption, and massive
• 5:47
  liquid cooling infrastructure. And for you listening
• 5:50
  right now running a business, this isn't just
• 5:52
  Wall Street trivia. The bill for the AI boom
• 5:54
  always comes due. Someone has to pay for all
• 5:57
  that copper and cooling. Exactly. When that much
• 6:00
  capital is required just to keep the servers
• 6:02
  running, it trickles down. It means your general
• 6:05
  cloud hosting costs are going to creep up. It
• 6:07
  means memory pricing spikes. The specific tier
• 6:10
  of enterprise software you rely on gets a little
• 6:13
  more expensive. You are looking at the source
• 6:15
  of those downstream effects. Precisely. All that
• 6:18
  physical silicon running incredibly hot around
• 6:20
  the clock is basically baking the next generation
• 6:23
  of intelligence. Which brings us directly from
• 6:26
  the hardware to the software it powers. Right.
• 6:28
  And this is where the dispatch shifts from financial
• 6:31
  metrics to what feels like, honestly, an existential
• 6:34
  shift in capability. OpenAI just shipped GPT
• 6:38
  -6, codenamed Astra. The big one. Yeah. And the
• 6:41
  generational jump in capability here is doing
• 6:43
  something entirely unprecedented to cybersecurity.
• 6:46
  Astra is the first model to cross the critical
• 6:48
  threshold under OpenAI's own internal preparedness
• 6:51
  framework. What's fascinating here is how that
• 6:53
  critical rating is actually quantified. The security
• 6:56
  metrics they publish represent a profound leap.
• 6:58
  I mean, they are wild. They are. On a standard
• 7:01
  industry benchmark called ExploitBench, Astra
• 7:04
  scored 100%. A perfect score. A perfect score.
• 7:07
  But the metric that should make every security
• 7:10
  professional pause is the harder test. They unleashed
• 7:13
  Astra on fresh vulnerabilities. Specifically,
• 7:17
  20 high -severity bugs found in Google Chrome's
• 7:21
  JavaScript engine just between June and August
• 7:24
  of this year. So totally new stuff. Exactly.
• 7:26
  These are novel, undocumented flaws. And Astra
• 7:30
  scored 39 % on that test. To put that in perspective,
• 7:33
  the previous model scored 5 .5%. That is a massive
• 7:37
  jump. Huge. And during the evaluation, OpenAI
• 7:39
  says the model actually found and exploited two
• 7:42
  previously unknown zero -day vulnerabilities
• 7:45
  entirely on its own. Just unbelievable. We are
• 7:48
  talking about it reading the underlying logic
• 7:49
  of a system and identifying a structural flaw
• 7:52
  that humans hadn't even categorized yet. It is
• 7:55
  not just regurgitating code it memorized from
• 7:57
  Stack Overflow anymore. It is demonstrating active
• 7:59
  reasoning. It is structural analysis. at machine
• 8:02
  speed. Now, to be clear, the consumer version
• 8:04
  of Astra that was actually shipped to the public
• 8:06
  has heavily restrictive guardrails. Oh, of course.
• 8:08
  Right. It will refuse to write proof of concept
• 8:11
  exploit code if you ask it to. But the ship product
• 8:13
  isn't really the point. The point is the capability
• 8:16
  exists in the wild now. And it represents a massive
• 8:19
  shift in the direction of travel. Historically,
• 8:22
  there's always been a window of time between
• 8:24
  a software vulnerability being discovered and
• 8:27
  a malicious actor figuring out how to weaponize
• 8:29
  it. Like a grace period. Exactly. Five years
• 8:33
  ago, that window was weeks. Lately, it has been
• 8:36
  days. This technology threatens to shrink that
• 8:39
  weaponization window to absolute zero. If an
• 8:42
  AI can analyze a patch, understand the vulnerability
• 8:45
  it fixes, and generate an exploit in seconds,
• 8:48
  well, human defenders simply cannot respond fast
• 8:51
  enough. Here's where it gets really interesting,
• 8:53
  though. Because on the exact same day they announced
• 8:55
  this terrifying new capability, OpenAI also announces
• 8:59
  a $1 billion subsidy for defensive AI tooling.
• 9:02
  Oh, the timing on that. Right. It is aimed at
• 9:05
  critical infrastructure, local governments, and
• 9:07
  open source maintainers. The details are a bit
• 9:10
  murky, you know. It looks like a six -month run
• 9:12
  prioritizing U .S. infrastructure. But the timing
• 9:16
  is just wild. OpenAI is heavily subsidizing cyber
• 9:20
  defense at the exact same time they're building
• 9:22
  the ultimate cyber offense tool. It's quite the
• 9:25
  paradox. It feels a bit like selling the locks
• 9:27
  and the lock picks at the same store, doesn't
• 9:29
  it? It really highlights the... unavoidable dual
• 9:31
  -use nature of artificial intelligence. OpenAI
• 9:35
  recognizes that by pushing the frontier of offensive
• 9:37
  capabilities forward, even if they keep the raw
• 9:40
  model internal, they are inherently destabilizing
• 9:43
  the broader security environment. Yeah, they
• 9:45
  have to know that. They do. The billion dollar
• 9:47
  subsidy is essentially a public acknowledgement
• 9:49
  that human defenders can no longer keep up with
• 9:52
  machine speed attacks unless they are equipped
• 9:54
  with machine speed defenses. But the irony here
• 9:56
  is that we are worrying about Astra weaponizing
• 9:59
  zero days in some hypothetical future. While
• 10:01
  right now, today, the infrastructure we built
• 10:04
  to host these AIs is practically leaving the
• 10:06
  door wide open. It is a mess out there. This
• 10:09
  shrinking threat window is happening today, and
• 10:12
  the first casualties are the very AI systems
• 10:14
  businesses just rushed to build. Yes, the so
• 10:17
  -called AI plumbing is leaking bad. Leaking everywhere.
• 10:21
  The dispatch highlights an urgent alert from
• 10:23
  CISA. That's the American Cybersecurity and Infrastructure
• 10:26
  Security Agency. In a single day, CISA added
• 10:30
  seven actively exploited vulnerabilities to their
• 10:33
  catalog. Seven in one day. And we should clarify.
• 10:37
  The CISA catalog isn't a list of theoretical
• 10:39
  bugs. If something lands on that list, it means
• 10:42
  someone is actively using it to break into systems
• 10:44
  right now in the real world. And four of those
• 10:47
  seven vulnerabilities were in AI and DevOps plumbing.
• 10:50
  We are talking about integration tools like Kestra,
• 10:53
  JFrog, Starlet, and an authentication flaw in
• 10:56
  Barry's Letellum proxy. A lot of people rely
• 10:59
  on those. They really do. And if you aren't familiar
• 11:01
  with these tools, they are essentially the glue.
• 11:03
  They move code around. They route your prompts
• 11:05
  to different language models. They handled the
• 11:06
  API keys. Microsoft observed attackers exploiting
• 11:10
  these specific flaws as early as June. And what
• 11:13
  the attackers were doing with that access is
• 11:15
  critical to understand. They weren't just, you
• 11:17
  know, asking the AI to write bad poetry for free.
• 11:20
  Right. If only it were that harmless. I know.
• 11:22
  They were deploying reverse shells, which essentially
• 11:25
  gives the hacker a hidden remote command line
• 11:28
  directly into the server. They were dropping
• 11:30
  crypto miners to steal computing power. And most
• 11:33
  dangerously, they were specifically targeting
• 11:35
  the database tables that hold provider API keys.
• 11:39
  If we look at how most businesses have operated
• 11:41
  over the last 18 months, there has been this
• 11:43
  mad rush to stand up AI gateways and proxies
• 11:48
  to connect company data to these large language
• 11:50
  models. But they are treating these AI gateways
• 11:53
  like simple, isolated internal websites. Which
• 11:56
  is a huge mistake. It is. Treating a highly connected
• 11:59
  AI gateway like a simple, isolated website is
• 12:02
  like leaving the master key to a bank vault under
• 12:04
  the front doormat because you assume nobody will
• 12:06
  look there. I mean, that gateway is connected
• 12:08
  to everything. Your billing API, your proprietary
• 12:10
  databases, your internal file systems. If we
• 12:13
  connect this to the bigger picture, Microsoft's
• 12:15
  explicit recommendation in this dispatch addresses
• 12:19
  exactly that architectural failure. Businesses
• 12:23
  must monitor these AI workloads according to
• 12:26
  their control plane role. The control plane.
• 12:29
  Yeah, the control plane is the nervous system
• 12:30
  of your network. Your AI gateway has the keys
• 12:34
  to your kingdom. you must treat it with the exact
• 12:37
  same security rigor, intense logging, and strict
• 12:41
  access controls that you would apply to a primary
• 12:43
  domain controller. Which makes total sense. But
• 12:46
  the dispatch takes a really bizarre turn right
• 12:49
  after this. Oh, yes, it does. It shifts from
• 12:51
  these hyper -complex, futuristic AI architecture
• 12:54
  problems down into the absolute mud of basic
• 12:58
  web plugins. It's jarring, honestly. It is. Because
• 13:01
  while we have frontier AI models finding novel
• 13:03
  zero days, Businesses are still getting completely
• 13:06
  compromised because they won't update their WordPress
• 13:08
  sites. It is the great paradox of modern cybersecurity.
• 13:11
  The most advanced multi -billion dollar threats
• 13:13
  share the exact same stage with the most basic
• 13:16
  human complacency. Let's look at this Elementor
• 13:19
  Pro vulnerability. If your business runs a WordPress
• 13:22
  site, you almost certainly know Elementor. It
• 13:25
  is a massive page builder plugin. Very popular.
• 13:28
  Super popular. Well, version 4 .2 .2 had a critical
• 13:33
  arbitrary file upload vulnerability. As of early
• 13:37
  September, roughly two -thirds of Elementor's
• 13:40
  10 million active installations were still running
• 13:42
  the vulnerable version. Two -thirds. That is
• 13:45
  millions of exposed sites. Millions. The security
• 13:49
  firm WordFence had to block over 190 ,000 exploit
• 13:52
  attempts almost instantly. The mechanism of the
• 13:55
  exploit itself is a perfect example of how a
• 13:57
  tiny logical flaw can completely compromise a
• 14:00
  system. Right, it's a flaw in the validation
• 14:02
  loop. Explain how that works. Okay, so normally
• 14:05
  when a user... files to a website, the system
• 14:08
  checks them one by one to ensure they're safe,
• 14:10
  checking that it is actually a standard JPEG
• 14:12
  image file, for example, and not some malicious
• 14:15
  executable script. Makes sense. But in this specific
• 14:17
  version of Elementor, if the validation loop
• 14:20
  hits an empty upload slot, it throws an error
• 14:22
  and just stops checking entirely. It skips validation
• 14:25
  for all the remaining files in the queue. Yes.
• 14:28
  So what does this all mean? It's like a bouncer
• 14:30
  at a club checking the ID of the first person
• 14:33
  in line, finding out the ID is completely blank,
• 14:36
  getting confused, throwing his hands up, and
• 14:39
  just letting the next 50 people walk right past
• 14:41
  him into the club, unchecked. It is a brilliant
• 14:44
  analogy. Because what the attacker does is submit
• 14:47
  an array of files. The first slot is deliberately
• 14:50
  left empty. The system hits it. gets confused,
• 14:53
  and stops the security check. And then the bad
• 14:56
  stuff gets through. Right, because the second
• 14:58
  slot contains a malicious PHP payload. Yeah.
• 15:00
  Because the validation stopped, that PHP file,
• 15:03
  which is essentially a script that the server
• 15:05
  will execute, gets written directly to the server's
• 15:07
  disk, completely unchecked. And once that PHP
• 15:09
  file is on the disk, it is game over. The attacker
• 15:12
  just navigates to that file's web address, the
• 15:14
  server executes it, and they have complete remote
• 15:17
  control over your entire website. The dispatch
• 15:19
  has a very clear, urgent action item here. If
• 15:22
  you use Elementor, you need to go check your
• 15:24
  web content slash upload slash Elementor slash
• 15:27
  forms directory. Yes, immediately. If there is
• 15:30
  a PHP file sitting in there, it is not a false
• 15:33
  positive. You have a backdoor and your site has
• 15:35
  been compromised. And this wasn't even the only
• 15:37
  mundane failure of the week. There was also a
• 15:40
  critical SQL injection flaw in Sangoma's Switchvox.
• 15:45
  Which is a VoIP phone system, right? Voice over
• 15:47
  IP, essentially the system routing all the office
• 15:49
  phone calls over the internet. Correct. And this
• 15:52
  flaw had a CVSS score of 9 .3 out of 10. Wow,
• 15:56
  9 .3. Yeah. The Common Vulnerability Scoring
• 16:00
  System ranks how severe and easy to exploit a
• 16:02
  bug is. A 9 .3 essentially means anyone with
• 16:06
  an internet connection and basic tools can likely
• 16:08
  break in without needing any specialized skills
• 16:11
  or prior access. So it's basically an open door.
• 16:14
  Pretty much, Sangoma patched this incredibly
• 16:16
  dangerous flaw in mid -July, but attackers were
• 16:19
  seen exploiting it in the wild at the end of
• 16:21
  August. That is a 47 -day window, 47 days where
• 16:25
  the patch was sitting there available, and most
• 16:26
  small businesses simply did not apply it, leaving
• 16:29
  their entire internal communication system exposed
• 16:31
  to the open internet. It happens all the time.
• 16:33
  And then, of course, to cap it all off, we have
• 16:36
  Google shipping their sixth zero -day Chrome
• 16:39
  patch of the entire year, fixing a flaw in the
• 16:42
  V8 JavaScript engine. This raises an important
• 16:45
  question. How do you apply a Chrome update? The
• 16:49
  browser literally does it automatically in the
• 16:51
  background. The fix is already downloaded and
• 16:54
  sitting there on most users' computers. So we
• 16:57
  have built these trillion -dollar intelligence
• 16:58
  systems, and the security of our network still
• 17:01
  comes down to... Did you close the window and
• 17:04
  turn it back on again? It forces us to confront
• 17:06
  the human element in all of this. The Chrome
• 17:08
  update is waiting for a user to simply close
• 17:10
  a browser window that they probably haven't closed
• 17:12
  since March because they don't want to lose their
• 17:14
  tabs. We're all guilty of tab hoarding. Oh, absolutely.
• 17:17
  But we can debate. billion dollar AI defense
• 17:21
  subsidies and autonomous threat hunting frameworks
• 17:23
  all day long. The single highest value, lowest
• 17:27
  effort security habit available to any office
• 17:29
  right now is just telling your staff to close
• 17:31
  their web browser completely once a week. It's
• 17:34
  so simple. It is. The attack surface of any business
• 17:37
  is ultimately managed by humans. Humans who are
• 17:40
  tired, distracted, or simply unaware that leaving
• 17:44
  a browser open for six months is a critical security
• 17:47
  risk. Let's quickly recap the journey we have
• 17:50
  been on today because it really is a wild ride.
• 17:52
  We zoomed way out to see the billions of dollars
• 17:56
  pouring into the physical foundation of AI through
• 17:59
  companies like Hugging Face and Broadcom's massive
• 18:01
  hardware build -out. The macro view. Right. Then
• 18:04
  we watched as AI crossed a critical, terrifying
• 18:07
  threshold with GPT. The overarching takeaway
• 18:26
  for anyone listening is that knowledge is only
• 18:28
  power if you actually apply it. Understanding
• 18:31
  the macro trends like the shrinking weaponization
• 18:33
  window is vital for your long -term strategy.
• 18:36
  But the micro actions like patching the VoIP
• 18:38
  system are what keep your business alive today.
• 18:41
  So for everyone listening, your homework is crystal
• 18:44
  clear. Go check your WordPress plugins right
• 18:46
  now and write down your AI model dependencies
• 18:49
  before the end of the day. But before we wrap
• 18:51
  up, I want to leave you with one final thought
• 18:53
  to mull over, building on everything we have
• 18:55
  explored in this dispatch. If we look at the
• 18:57
  trajectory we have mapped out today, as AI capabilities
• 19:01
  shrink that vulnerability window to near zero,
• 19:04
  And as models like GPT -6 begin to find and exploit
• 19:07
  novel zero days entirely on their own, are we
• 19:10
  rapidly approaching a threshold where humans
• 19:12
  are completely removed from the cybersecurity
• 19:15
  loop? That is a terrifying thought. We might
• 19:18
  soon be facing a reality where AI perpetually
• 19:20
  attacks and patches AI infrastructure at speeds
• 19:23
  that human cognition cannot even perceive. We
• 19:26
  could be looking at the dawn of cyber warfare
• 19:28
  as a machine -only domain, where we are just
• 19:30
  the bystanders, entirely reliant on the structural
• 19:33
  integrity of the code. that is a staggering thought
• 19:35
  a silent invisible war happening millions of
• 19:39
  times a second in the background just to keep
• 19:41
  our browser tabs from crashing thank you all
• 19:43
  for joining us on this deep dive into the architecture
• 19:45
  of our very weird very fast moving future good
• 19:49
  luck out there in the digital wild and seriously
• 19:51
  go restart your browser we'll catch you next
• 19:53
  time