Latest / Tech Talks With Kinsoft / Last Week in Tech – Gemini 3.1 and Claude 4.6 Land, OpenAI's Agent Hire, and the Figure Fintech Breach

Transcript

Last Week in Tech – Gemini 3.1 and Claude 4.6 Land, OpenAI's Agent Hire, and the Figure Fintech Breach · Tech Talks With Kinsoft

• 0:00
  Welcome to the deep dive. So our mission today
• 0:02
  is to unpack this. Honestly, it's a completely
• 0:06
  wild juxtaposition that we found buried inside
• 0:10
  a late February 2026 industry brief. Yeah. From
• 0:14
  Kinsoft Tech Talks. Right. Exactly. From Kinsoft.
• 0:17
  And we're basically looking at this snapshot
• 0:19
  of just a single week in the tech sector. And
• 0:21
  the contrast is, I mean, it's genuinely mind
• 0:24
  bending. It really is. Because on one side of
• 0:26
  the ledger, you've got these tech giants, you
• 0:29
  know, rolling out hyper advanced, totally autonomous
• 0:31
  AI systems that are like verging on science fiction.
• 0:34
  Oh, absolutely. Stuff we couldn't imagine five
• 0:36
  years ago. Right. But then on the exact same
• 0:38
  page of this report, we're seeing these massive
• 0:40
  catastrophic data breaches causing just absolute
• 0:43
  chaos. And they're all triggered by vulnerabilities
• 0:46
  so basic they barely even qualify as hacks. Yeah.
• 0:49
  And I think it's a critical dynamic for you to
• 0:51
  understand right now as you're listening to this.
• 0:53
  Because, I mean, whether you're an enterprise
• 0:56
  engineering lead trying to architect a secure
• 0:59
  AI rollout for your company or... you know, just
• 1:03
  a regular user trying to ensure your personal
• 1:05
  financial footprint isn't scraped and sold on
• 1:07
  the dark web. Which is all of us, basically.
• 1:09
  Right, exactly. Understanding how these two extremes
• 1:12
  collide, it's probably the most important navigation
• 1:15
  skill you can have in the current tech landscape.
• 1:18
  For sure. And to put this into perspective, imagine
• 1:22
  you're in your garage building a car. Right.
• 1:24
  Okay, I'm with you. And you somehow manage to
• 1:27
  acquire this million horsepower, state -of -the
• 1:29
  -art experimental jet engine. and you strap it
• 1:33
  directly onto the chassis bolt it right on bolt
• 1:35
  it right on and the engineering is flawless right
• 1:38
  it can take you anywhere instantly but then you
• 1:41
  step back to admire your work you look at the
• 1:42
  doors and you realize the locks on this magnificent
• 1:45
  machine can still be picked from the outside
• 1:47
  with like a rusty paper clip the jet engine next
• 1:51
  to the paper clip lock it's such a perfect analogy
• 1:53
  crazy right It is. I mean, you have the absolute
• 1:56
  pinnacle of machine reasoning layered directly
• 2:00
  on top of these glaring, primitive structural
• 2:02
  vulnerabilities. Yeah. So let's start with the
• 2:05
  jet engine side of the equation first, because
• 2:07
  late February was, I mean, it was an absolute
• 2:10
  arms race. It was a wild week. Right. The Kinsoft
• 2:12
  brief highlights these three massive updates
• 2:16
  hitting the industry. practically at the exact
• 2:18
  same time. We've got Google, Anthropic, and OpenAI.
• 2:22
  Let's start with Google dropping Gemini 3 .1
• 2:25
  Pro in preview. Yeah, that was huge. So a 0 .1
• 2:28
  release in machine learning usually implies like
• 2:31
  a major architectural flex, right? Not just a
• 2:34
  simple fine -tuning pass. So what is actually
• 2:36
  happening under the hood there? Well, the 0 .1
• 2:38
  designation here signals a really significant
• 2:41
  evolution in how the model routes information.
• 2:43
  Okay. They aren't just scaling up raw compute.
• 2:46
  They're refining the underlying architectures,
• 2:49
  likely leaning even heavier into what we call
• 2:52
  advanced sparse expert models. Sparse expert
• 2:55
  models. So what does that mean in plain English?
• 2:57
  Right. So instead of activating the entire neural
• 3:00
  network for every single query, which is computationally
• 3:04
  ruinous and super expensive. Like turning on
• 3:06
  all the lights in a skyscraper just to read a
• 3:08
  book in one room. Exactly. Great way to put it.
• 3:11
  Instead of doing that, the system dynamically
• 3:13
  routes your prompt to highly specialized subnetworks.
• 3:17
  I see. And the result is what they're calling
• 3:19
  class -leading benchmark scores achieved with
• 3:22
  just far better inference efficiency. Wow. And
• 3:26
  the headline feature they attach to this efficiency
• 3:28
  is that one million token context window. Yes,
• 3:32
  the million tokens. Which is wild. And we'll
• 3:34
  get back to why that specific number is actually
• 3:36
  so dangerous in a minute. But first, let's look
• 3:39
  at how Anthropic responded. Right on Google's
• 3:41
  heels, they shipped Claude Sonnet 4 .6. Yeah,
• 3:44
  they didn't waste any time. No time at all. And
• 3:46
  Sonnet is historically their middle tier. model
• 3:49
  right the one that optimizes for speed and cost
• 3:51
  exactly but with 4 .6 they pushed this massive
• 3:54
  upgrade specifically targeting coding capabilities
• 3:57
  and these long multi -step tasks and they kept
• 4:00
  the price exactly the same which is the craziest
• 4:03
  part to me The mechanism driving that price to
• 4:06
  performance ratio is fascinating. How do they
• 4:08
  even do that without burning cash? Well, to boost
• 4:11
  reasoning on those long, multi -step tasks without
• 4:16
  raising the API cost, you have to drastically
• 4:19
  optimize the working memory. It's what's called
• 4:23
  the KV cache. The KV cache, okay. Yeah. So Anthropic
• 4:26
  basically managed to compress how the model stores
• 4:30
  the history of the conversation. Oh, I see. So
• 4:32
  this allows it to maintain a really deep understanding
• 4:35
  of complex code bases without requiring this
• 4:39
  massive, expensive spike in compute for every
• 4:42
  single new turn of the conversation. Okay, that
• 4:44
  makes sense. So they're just getting way more
• 4:45
  efficient with the memory they already have.
• 4:47
  Exactly. And then, okay, we have OpenAI. And
• 4:49
  the Kinsoft brief points out that they take a
• 4:51
  totally different angle that week. Instead of
• 4:53
  dropping a massive new model, They made a really
• 4:56
  strategic acquisition of human talent. Yeah,
• 4:58
  this is a massive signal to the market. Right.
• 5:00
  They hired the creator of this massively popular
• 5:03
  open source project to lead their push into what
• 5:07
  they explicitly called next generation personal
• 5:10
  agents. Yeah. And the focus on the open source
• 5:12
  community here is the absolute tell. How so?
• 5:16
  Well. Over the last two years, while the big
• 5:19
  labs were focused on building the underlying
• 5:20
  foundational models, like the raw brains, the
• 5:24
  open source community was busy perfecting the
• 5:26
  orchestration layer. Okay, the orchestration
• 5:28
  layer. Meaning what exactly? They were the ones
• 5:30
  figuring out the scaffolding. Like how to take
• 5:32
  a raw language model and actually give it tools,
• 5:35
  give it memory, and give it agency. Oh, right.
• 5:37
  So it can actually do things instead of just
• 5:39
  talking to you. Yes. Open AI bringing in that
• 5:42
  specific talent. signals a huge structural pivot.
• 5:45
  They realize that the raw intelligence of the
• 5:48
  model is no longer the primary bottleneck. Right.
• 5:50
  The models are already smart enough. Exactly.
• 5:52
  The bottleneck now is the framework that allows
• 5:54
  that model to interact with the external world
• 5:57
  seamlessly. Which brings us to the giant neon
• 5:59
  sign flashing over all three of these updates,
• 6:01
  which is the word agent. Yep. The era of the
• 6:05
  agent. The Kinsoft brief is just adamant about
• 6:07
  this. The era of the passive assistant is completely
• 6:10
  over. We are now in the era of the agent. So
• 6:13
  for someone who has spent the last few years
• 6:15
  treating AI like an incredibly smart answering
• 6:19
  machine, what's the actual mechanical shift happening
• 6:22
  here? So think about an assistant. An assistant
• 6:24
  operates in a closed loop initiated by the user.
• 6:27
  Okay. You send a prompt. It retrieves or generates
• 6:31
  data, it outputs text, and then it just stops.
• 6:33
  It waits for you to drive. Like a search engine
• 6:36
  on steroids. Exactly. But an agent is asynchronous
• 6:39
  and goal -oriented. You give it an objective
• 6:41
  like... audit our vendor's software spend for
• 6:45
  Q3, and highlight discrepancies. Wow, okay. Right,
• 6:47
  and it breaks that big goal into smaller tasks
• 6:49
  on its own. It generates a plan, writes a script
• 6:52
  to query the database, reads the output, realizes
• 6:55
  a data field is missing. And it fixes it. Yes.
• 6:57
  It accesses a secondary API to pull the missing
• 7:00
  data, compiles the report, and then emails it
• 7:03
  to your team. It executes an entire operational
• 7:05
  loop all on its own. That is just wild. And this
• 7:08
  connects directly back to my earlier question
• 7:10
  about Google's 1 million token context window,
• 7:14
  because we hear these massive token counts thrown
• 7:16
  around constantly. Right. It's the favorite marketing
• 7:19
  metric right now. Exactly. But I want to look
• 7:21
  at the actual utility of this for you, the listener.
• 7:24
  What does a million tokens allow an AI agent
• 7:28
  to do that a smaller context window prohibits?
• 7:31
  That's the key question. Because, I mean, if
• 7:33
  an assistant just needs to summarize a thick
• 7:35
  PDF, sure. A big window is great, but for an
• 7:38
  active agent, is it just about reading more text
• 7:41
  or is this about maintaining a chain of actions
• 7:43
  without losing the plot? It is entirely about
• 7:46
  maintaining state over a prolonged period. So
• 7:49
  think of the context window not as a bookshelf
• 7:52
  where it stores information, but as the physical
• 7:54
  desk where the agent is actively working. Okay,
• 7:56
  the physical desk. And a million tokens means...
• 7:59
  A massive desk. A desk roughly the size of a
• 8:01
  football field. Oh. Yeah. If an agent is executing
• 8:04
  a really complex task that takes, say, four hours.
• 8:08
  It needs to keep the original instructions, the
• 8:11
  API documentation for the tools it's using, the
• 8:13
  intermediate code it just wrote, and the error
• 8:16
  logs it just generated all spread out on the
• 8:18
  desk at the exact same time. Right. It needs
• 8:20
  to see all of it at once. Exactly. If the context
• 8:22
  window is small, the desk is tiny. The agent
• 8:25
  has to constantly throw old documents in the
• 8:27
  trash just to make room for the new ones. And
• 8:30
  then suddenly it forgets why it was querying
• 8:32
  the database in the first place. Precisely. It
• 8:35
  suffers from context amnesia. Context amnesia.
• 8:38
  That's a great term for it. Yeah, it starts hallucinating
• 8:40
  variables or repeating the exact same error loop
• 8:43
  because it dropped the initial instructions out
• 8:45
  of its working memory. I've actually seen that
• 8:47
  happen with older models. It just gets stuck
• 8:49
  in a loop. We all have. So a one million token
• 8:52
  window means the agent can hold the entire sprawling
• 8:55
  history of a massive multi -step task constantly
• 8:58
  accessible in its active mind. It can execute
• 9:02
  sophisticated workflows over days without losing
• 9:04
  its train of thought. Which is incredible. We
• 9:07
  are building autonomous systems capable of roaming
• 9:10
  through our digital infrastructure, making decisions,
• 9:13
  writing code, moving files. I mean, this is the
• 9:16
  million horsepower jet engine. It really is.
• 9:18
  But here is the hard pivot in the Kinsoft briefing.
• 9:21
  And this is where things get incredibly precarious.
• 9:24
  Yeah, this is the scary part. Right. Because
• 9:27
  if an agent is going to execute these complex
• 9:29
  tasks for you, it requires access. It needs read
• 9:33
  and write permissions to your sensitive company
• 9:35
  databases. It needs access to your personal financial
• 9:38
  accounts. It requires the keys to the kingdom.
• 9:40
  You are literally delegating authentication to
• 9:43
  a machine. Delegating authentication, yeah. And
• 9:46
  as Debris points out, if you're handing over
• 9:48
  the keys to highly capable autonomous systems,
• 9:51
  you have to look very closely at the actual structural
• 9:54
  integrity of the locks. drags us right into the
• 9:56
  second half of this briefing. And I'm just going
• 9:58
  to say it. This is the absolute basement of digital
• 10:01
  security. It's grim. It really is. Because while
• 10:05
  the major AI labs were shipping these world -changing
• 10:07
  updates, the exact same week was dominated by
• 10:10
  a series of catastrophic data breaches. Right.
• 10:13
  And the contrast is just stark because the threats
• 10:16
  described here are shockingly unsophisticated.
• 10:19
  Like, let's look at the first case study Kinsoft
• 10:22
  highlights. This is Figure Technology, a major
• 10:25
  fintech lender. Yeah. So Figure Technology disclosed
• 10:28
  a breach that exposed the personal data, meaning
• 10:30
  names, addresses, dates of birth. of roughly
• 10:34
  967 ,000 customers. Almost a million people.
• 10:38
  Almost a million people. Now, when we hear about
• 10:41
  a fintech platform getting breached, the Hollywood
• 10:43
  assumption is, you know, a highly sophisticated
• 10:45
  zero -day exploit bypassing layers of advanced
• 10:47
  cryptography. Right, like some elite hacking
• 10:49
  syndicate tunneling through the firewall with
• 10:52
  custom malware, green code scrolling down the
• 10:54
  screen. Exactly. But the reality is just... painfully
• 10:57
  mundane. The breach was executed by the shiny
• 11:00
  hunters extortion crew and their primary vector
• 11:02
  wasn't a zero day exploit. It was social engineering
• 11:05
  and compromised credentials. OK. So this is basically
• 11:08
  the digital equivalent of a con artist. Essentially
• 11:10
  yes. They aren't cracking the vault. They're
• 11:13
  just wearing a high vast vest holding a clipboard
• 11:16
  and walking up to the front door and someone
• 11:18
  inside the building literally holds the door
• 11:20
  open for them. That is exactly what happens.
• 11:22
  But wait how does this actually work at scale.
• 11:25
  Because if a company like tech has modern security
• 11:27
  in place how does a crew like shiny hunters actually
• 11:30
  bypass the authentication don't they have safeguards
• 11:33
  they do but attackers rely on mechanisms like
• 11:37
  adversary in the middle or ATM architecture ATM
• 11:41
  okay break that down for me so they don't just
• 11:43
  ask for a password they set up a reverse proxy
• 11:46
  server when an employee is tricked into clicking
• 11:49
  a phishing link say an urgent text message that
• 11:53
  looks exactly like it's from their IT help desk.
• 11:55
  Which happens all the time. All the time. So
• 11:57
  the employee clicks it and they are routed through
• 11:59
  this proxy. The employee types in their credentials
• 12:01
  and the proxy intercepts them. Oh, man. But more
• 12:05
  importantly, when the actual corporate server
• 12:07
  sends back the session cookie to validate the
• 12:09
  login, the proxy steals that session cookie too.
• 12:13
  Wait, so they bypassed multi -factor authentication
• 12:15
  completely? Like the system thinks the hacker
• 12:18
  is the verified employee who just successfully
• 12:20
  completed the MFA challenge? Exactly. It exploits
• 12:24
  the human psychological response to urgency.
• 12:27
  capturing the authenticated session without ever
• 12:30
  needing to break the underlying encryption. That
• 12:32
  is terrifying. It is. And we see this exact same
• 12:35
  reliance on human vulnerability in the second
• 12:38
  major breach, the brief highlights, which targeted
• 12:41
  France's national bank account registry. Right.
• 12:44
  And this is a nation state level target. Data
• 12:46
  tied to roughly 1 .2 million accounts was accessed.
• 12:49
  And again, this wasn't some sophisticated cyber
• 12:52
  weapon taking down the mainframe, right? Nope.
• 12:54
  The registry was compromised using a single government
• 12:56
  employee's stolen credentials. A single employee.
• 13:00
  I mean, how does one compromised login expose
• 13:03
  over a million bank accounts? That implies an
• 13:05
  incredibly flat network. Like if I steal a bank
• 13:08
  teller's badge, it shouldn't give me access to
• 13:09
  every safe deposit box in the country. Yeah,
• 13:12
  that is the core architectural failure right
• 13:14
  there. It is a total lack of data compartmentalization
• 13:18
  and a failure to implement a zero trust architecture.
• 13:21
  Zero trust. We hear that buzzword a lot. Yeah.
• 13:24
  In a legacy system, once you pass the perimeter
• 13:26
  like, once you log in, the system just assumes
• 13:28
  you are trusted and grants broad lateral access.
• 13:32
  So you can just roam around the network. Exactly.
• 13:34
  Zero trust, by contrast, operates on the assumption
• 13:37
  that the network is always hostile. It requires
• 13:41
  continuous verification. So it's always asking,
• 13:43
  are you sure you're allowed to be here? Yes.
• 13:45
  If a mid -level government employee suddenly
• 13:48
  attempts to export a database of 1 .2 million
• 13:51
  accounts, a zero -trust system flags that anomaly
• 13:54
  and requires secondary out -of -band authorization.
• 13:57
  Right. It stops them in their tracks. Exactly.
• 13:59
  The fact that the data was exfiltrated using
• 14:02
  basic credentials shows that once the attackers
• 14:04
  were inside, there were absolutely no internal
• 14:07
  blast doors to smack them. Wow. So we have two
• 14:10
  massive breaches, nearly two million people exposed,
• 14:13
  all because of the paperclip lock, right? Someone
• 14:15
  handing over their password. But the Kinsoft
• 14:18
  brief includes a third incident from that same
• 14:20
  week, PayPal. Ah, yes. The PayPal incident. And
• 14:24
  this one is different, right? Because it doesn't
• 14:26
  involve a stolen credential or an extortion crew
• 14:29
  like shiny hunters. Right. So PayPal disclosed
• 14:31
  that a software bug in their loan application
• 14:33
  flow quietly exposed personal data for months.
• 14:37
  There was no active manipulation of an employee
• 14:40
  at all. So how does a bug like that just sit
• 14:43
  undetected on one of the largest payment platforms
• 14:45
  on Earth? I mean, for months. Well, while they
• 14:48
  didn't release the exact technical postmortem,
• 14:51
  issues like this almost always stem from... bella
• 14:53
  -lay vulnerabilities. Bella -lay. Yeah, broken
• 14:56
  object -level authorization. Okay, broken object
• 14:58
  -level authorization. What is that? So in modern
• 15:00
  applications, your front -end web browser constantly
• 15:02
  communicates with back -end servers through APIs,
• 15:05
  right? Yeah. A bowl of vulnerability happens
• 15:06
  when the API endpoint doesn't properly verify
• 15:11
  if the user requesting the data actually has
• 15:14
  the authorization to view that specific data.
• 15:18
  Wait, so the system checks that you're logged
• 15:20
  into PayPal, but it forgets to check if you actually
• 15:23
  own the specific loan application you're asking
• 15:25
  to see. That is the exact mechanism. Yeah, seriously.
• 15:28
  Yes. If a developer just forgets to write the
• 15:31
  specific line of code that matches the user ID
• 15:33
  to the document ID on the back end, a user can
• 15:36
  simply modify the application number in the URL
• 15:39
  and the server will happily serve up someone
• 15:42
  else's sensitive financial data. Because it's
• 15:44
  an exposed endpoint. Exactly. So it's the equivalent
• 15:46
  of like building a bank vault with three foot
• 15:48
  thick solid steel walls, but forgetting to actually
• 15:51
  install the hinges on the door. It just falls
• 15:54
  over if you lean on it. That is. unfortunately
• 15:56
  a very accurate way to describe it man okay so
• 15:59
  we have this kinsoft briefing advising the entire
• 16:01
  tech industry and they are analyzing ai agents
• 16:04
  that can hold a million tokens in working memory
• 16:07
  to execute code while simultaneously watching
• 16:10
  massive databases get cracked because a developer
• 16:12
  missed an api authorization check or a government
• 16:16
  employee clicked a bad link on their lunch break
• 16:18
  it's quite the paradox it's insane Given this
• 16:22
  reality, what is their actual advice? How does
• 16:25
  an organization defend itself in a threat landscape
• 16:27
  that relies on such low -tech methods? Well,
• 16:30
  the briefing's advice is incredibly grounded,
• 16:33
  actually. Because the vulnerabilities exploit
• 16:35
  basic structural and human flaws, the defense
• 16:38
  doesn't require futuristic next -generation technology.
• 16:41
  Oh, really? Yeah. The heavy lifting comes down
• 16:43
  to basic digital hygiene, specifically multi
• 16:46
  -factor authentication and the rigorous handling
• 16:48
  and compartmentalization of personal data. Okay,
• 16:51
  I have to push back hard on this. Every single
• 16:53
  person listening right now knows what MFA is.
• 16:55
  Of course they do. Every IT department on earth
• 16:57
  beats the drum of basic hygiene and zero trust.
• 17:01
  If MFA and data compartmentalization do the heavy
• 17:03
  lifting, and the entire industry already knows
• 17:05
  this, Why are massive organizations like a national
• 17:09
  government registry or a giant fintech lender
• 17:12
  still leaving themselves completely open? Are
• 17:14
  we just structurally lazy? It isn't laziness,
• 17:17
  actually. It is the inherent friction between
• 17:20
  security and enterprise velocity. Friction versus
• 17:23
  velocity. Okay. Security, by its very nature,
• 17:26
  introduces roadblocks. Multi -factor authentication
• 17:29
  is friction. Compartmentalizing data so an employee
• 17:33
  has to submit an IT ticket just to access a database
• 17:36
  that is friction. Yeah, and nobody likes filing
• 17:38
  IT tickets. Right. And in fast -paced environments,
• 17:41
  whether you're a booming fintech startup rushing
• 17:43
  to deploy new features or a sprawling government
• 17:46
  agency handling thousands of requests a day,
• 17:48
  friction is viewed as the absolute enemy of productivity.
• 17:51
  People just want to get their jobs done without
• 17:53
  authenticating for the 12th time before lunch.
• 17:55
  And extortion crews weaponize that exact fatigue.
• 17:57
  They execute what's called MFA prompt bombing.
• 17:59
  Oh, I've heard of this. Yeah, they send continuous
• 18:01
  authentication requests to an employee's phone
• 18:03
  until the employee gets so annoyed they finally
• 18:05
  just click a... prove to make the notification
• 18:08
  go away so they can get back to work. Which is
• 18:10
  just classic human psychology. True security
• 18:13
  against these attacks requires hardware keys
• 18:15
  like FIDO2 compliant physical tokens, which eliminate
• 18:19
  the vulnerability of those ATM proxies we talked
• 18:22
  about. Oh, right. Because the physical key proves
• 18:24
  you're really there. Exactly. But deploying physical
• 18:28
  hardware keys to thousands of employees is expensive
• 18:31
  and logistically complex. So organizations settle
• 18:34
  for less secure app -based MFA. They accept a
• 18:37
  lower security threshold in exchange for higher
• 18:40
  operational velocity. Which makes human error
• 18:42
  the persistent, unpatchable weak link in the
• 18:46
  entire infrastructure. I mean, you can have the
• 18:48
  most sophisticated encryption algorithms on the
• 18:50
  planet, but if a fatigued human has the authority
• 18:52
  to bypass them, the human becomes the attack
• 18:55
  vector. Unfortunately, yes. And this brings us
• 18:57
  to the Kinsoft sign -off, which is remarkably
• 18:59
  succinct. Their final advice to the industry
• 19:02
  is simply to stay patched, stay skeptical. Yeah.
• 19:06
  Stay skeptical is basically the psychological
• 19:08
  firewall. It is the only immediate defense. against
• 19:11
  social engineering. Just trusting your gut when
• 19:13
  something feels off. Well, more than that, it's
• 19:16
  the conscious insertion of friction back into
• 19:18
  the process, like pausing to verify out of band
• 19:22
  if the IT department is actually requesting a
• 19:25
  password reset. It's demanding to see the ID
• 19:28
  badge of the guy in the high -vis vest instead
• 19:30
  of just assuming he belongs in the server room.
• 19:32
  Exactly. And staying patched directly addresses
• 19:35
  the PayPal scenario. It is an acknowledgment
• 19:38
  that modern applications are incredibly complex
• 19:41
  and things like BOLA vulnerabilities and exposed
• 19:44
  API endpoints will occur. Because humans write
• 19:46
  the code and humans make typos. Right. So it
• 19:49
  requires automated, rigorous auditing of code
• 19:52
  bases to catch those structural errors before
• 19:54
  they sit exposed to the public Internet for months
• 19:56
  on end. All right. Let's pull this all together.
• 19:58
  Our mission today was to unpack this juxtaposition
• 20:01
  in the Kinsoft tech brief. And we're looking
• 20:03
  at two entirely different realities colliding
• 20:05
  here. Yes. Head on. On one side, we have this
• 20:08
  AI arms race pushing the boundaries of machine
• 20:10
  capability. You've got Google deploying Gemini
• 20:13
  3 .1 Pro with Moe architecture and a million
• 20:16
  token memory desk. Football field desk. Right.
• 20:19
  Anthropic optimizing the KV cache in CloudSonic
• 20:22
  4 .6 for complex coding tasks. OpenAI pivoting
• 20:26
  their entire corporate structure to build next
• 20:28
  -gen personal agents. We are officially in an
• 20:31
  era where AI doesn't just retrieve information,
• 20:34
  it actively navigates our digital infrastructure
• 20:36
  to execute prolonged tasks. The capabilities
• 20:39
  are scaling exponentially, and we are eagerly
• 20:42
  handing these agents the keys to our most sensitive
• 20:44
  data to unlock that productivity. But then, on
• 20:47
  the exact same day, we are reminded of the stark
• 20:49
  reality of how we protect the infrastructure
• 20:51
  those AI agents are interacting with. Figure
• 20:54
  technology and the French bank registry were
• 20:56
  breached not by supercomputers breaking cryptography,
• 20:59
  but by stolen passwords and proxy servers exploiting
• 21:01
  human psychology. PayPal exposed data because
• 21:04
  of a missed authorization check on a backend
• 21:06
  API. As we build AI agents capable of incredibly
• 21:10
  complex reasoning, the security perimeter defending
• 21:13
  our digital world remains entirely dependent
• 21:15
  on humans not falling for a phishing text. We
• 21:18
  are trusting highly advanced automated systems
• 21:21
  to handle our data, while the perimeter defending
• 21:24
  that data is guarded by human fatigue and basic
• 21:27
  software configuration errors. It fundamentally
• 21:30
  changes how you look at that million horsepower
• 21:32
  jet engine strapped to the chassis. I mean, if
• 21:34
  we are going to travel this fast, we urgently
• 21:36
  need to reengineer the door locks. We do. And,
• 21:40
  you know, following the logic of these two trends
• 21:43
  colliding actually raises a deeply provocative
• 21:45
  question regarding the future of enterprise architecture.
• 21:47
  Oh, where does this inevitably lead? Well, consider
• 21:51
  the core vulnerability here. It's human error,
• 21:53
  right? Yeah. Always. A government employee authorizing
• 21:56
  a malicious proxy because they were suffering
• 21:59
  from alert fatigue. A rushed programmer missing
• 22:02
  a Bidule vulnerability in a massive code base.
• 22:06
  Humans are consistently the weakest link. Because
• 22:08
  we get tired, we take shortcuts to maintain velocity,
• 22:11
  and our attention spans degrade. We're only human.
• 22:14
  Exactly. But simultaneously, we are perfecting
• 22:18
  autonomous AI agents that never suffer from alert
• 22:20
  fatigue. They cannot be socially engineered over
• 22:23
  a phone call. And they can analyze millions of
• 22:25
  tokens of code to identify unauthenticated API
• 22:28
  endpoints in seconds. Oh, wow. I see where you're
• 22:31
  going with this. We are building the exact system
• 22:33
  needed to counter the human flaws. Think about
• 22:36
  it. If human friction is the vulnerability, And
• 22:39
  these new AI agents are the solution for executing
• 22:41
  complex digital governance flawlessly. The ultimate
• 22:44
  end state of this arms race might be using these
• 22:47
  exact agents to lock humans out of our own security
• 22:49
  infrastructure entirely. Wait, lock us out entirely?
• 22:52
  Yes. If the AI agent is smarter than the extortion
• 22:55
  crew, and it's more meticulous than the human
• 22:57
  developer, the most secure architecture moving
• 23:00
  forward might be one where humans are no longer
• 23:02
  permitted to hold the keys at all. That is wild.
• 23:06
  Access controls. Database routing, code audits,
• 23:09
  all of it could be entirely governed by agentic
• 23:12
  AI, removing the human bottleneck and the human
• 23:15
  vulnerability completely. So to fix the rusty
• 23:18
  paperclip lock, we just take the doors off entirely,
• 23:20
  weld the chassis shut, and let the jet engine
• 23:23
  drive itself. We build the perfect machine, and
• 23:26
  then we have to remove ourselves from the loop
• 23:27
  so we don't accidentally compromise it. That
• 23:30
  is, man, that is a heavy concept to leave you
• 23:32
  with. But an essential one, as we navigate this
• 23:35
  wild transition from passive assistants to autonomous
• 23:38
  agents, stay patched, stay skeptical, and pay
• 23:41
  very close attention to who or what actually
• 23:44
  holds the keys to your data.