Latest / Tech Talks With Kinsoft / Last Week in Tech – Anthropic's $900B Round, AI Guardrails Stripped in Minutes, and Trellix's Source-Code Breach

Transcript

Last Week in Tech – Anthropic's $900B Round, AI Guardrails Stripped in Minutes, and Trellix's Source-Code Breach · Tech Talks With Kinsoft

• 0:00
  Welcome to the Deep Dive. Today, we are exploring
• 0:03
  what I can really only call this, you know, wild,
• 0:06
  dizzying paradox in the current tech landscape.
• 0:09
  Yeah, it's a lot to wrap your head around, honestly.
• 0:11
  It really is. So we have a single, just densely
• 0:14
  packed source for you today. It's the June 1st,
• 0:16
  2026 edition of Kinsoft Tech Talks. And the report
• 0:20
  is called The AI Economy and Security Frontiers.
• 0:24
  Right, which sounds very formal, but the contents
• 0:27
  are just... They're wild. Completely. And our
• 0:30
  mission for this deep dive is to basically navigate
• 0:33
  this incredibly stark contrast. Like we are going
• 0:36
  to take you from the staggering just money printing
• 0:39
  heights of this current AI boom right down to
• 0:42
  the, frankly, alarming vulnerabilities in the
• 0:45
  cybersecurity foundation that's supposedly holding
• 0:48
  it all up. And that is really the crucial dynamic
• 0:50
  to understand right now. My goal today is to
• 0:52
  help you, the listener, connect the dots between
• 0:54
  that. you know, unprecedented economic acceleration
• 0:57
  and the actual security fragility underneath
• 1:00
  it all. Because it feels like two completely
• 1:01
  different realities, right? Exactly. Because
• 1:04
  when you read the financial headlines, you see
• 1:06
  one very specific version of the future. But
• 1:09
  then, well, when you look at the infrastructure
• 1:11
  and the daily security realities, a completely
• 1:14
  different, much more precarious picture emerges.
• 1:16
  I have to say, just looking at the sheer scale
• 1:19
  of the information in this Kinsoft report. It
• 1:22
  is a lot to process. I mean, the numbers, the
• 1:25
  speed of development, the geopolitical stakes.
• 1:28
  It really requires a total reframing of how we
• 1:31
  think about technology. It does. It absolutely
• 1:34
  does. And that is exactly why we are breaking
• 1:36
  it down today, because understanding this landscape,
• 1:38
  it isn't just for like. tech insiders anymore.
• 1:41
  You know, it dictates the environment you are
• 1:44
  going to be operating in, regardless of what
• 1:45
  industry you're actually in. Right. OK, let's
• 1:47
  unpack this, because the momentum of the tech
• 1:49
  world right now is, I mean, fundamentally staggering.
• 1:53
  Yeah, it's moving at a pace we haven't really
• 1:54
  seen before. So the report kicks off with Anthropic,
• 1:57
  the company behind the Claude AI assistant. And
• 2:00
  according to the source, they are closing this
• 2:03
  30 billion dollar funding round. 30 billion.
• 2:06
  Which is just a massive injection of capital.
• 2:08
  And that puts their valuation north of $900 billion.
• 2:13
  $900 billion. Yeah, a valuation that basically
• 2:15
  places them among the most valuable entities
• 2:18
  in human history. And they achieved it in like
• 2:23
  a fraction of the time it took legacy tech giants
• 2:25
  to reach similar heights. And it's not just...
• 2:28
  you know, purely speculative value based on some
• 2:31
  future promises. The Kinsoff report notes that
• 2:34
  Anthropic is actually on track for its first
• 2:36
  ever quarterly operating profit. Right. They
• 2:39
  are projecting nearly $11 billion in revenue
• 2:42
  for the second quarter alone. $11 billion in
• 2:44
  a quarter. And at the exact same time, OpenAI
• 2:47
  is reportedly preparing a confidential filing
• 2:49
  to go public. I mean, the amount of capital just
• 2:52
  swirling around this space right now is immense.
• 2:54
  It really is. But then, you know, you look at
• 2:56
  the expense side of the ledger. Oh, man. Yeah.
• 2:58
  Which is arguably the most telling part of this
• 3:01
  entirely economic equation, really. Because in
• 3:03
  SpaceX's recent filings, it was revealed that
• 3:06
  Anthropic is paying SpaceX over a billion dollars
• 3:09
  a month for GPU compute. A billion dollars. Every
• 3:12
  single month. It's just it's hard to fathom.
• 3:15
  I want to offer an analogy here because I keep
• 3:17
  thinking about like the Industrial Revolution.
• 3:20
  That's a good comparison. Yeah. Right. You look
• 3:22
  back at that era and you had these massive tycoons
• 3:25
  building the giant, highly profitable factories
• 3:27
  that just fundamentally changed society. The
• 3:30
  Carnegie's, the Ford's. Exactly. And right now,
• 3:33
  Anthropic and OpenAI, they are building the modern
• 3:37
  equivalent of those massive factories. But SpaceX,
• 3:40
  SpaceX. is the one selling them the coal to keep
• 3:43
  the furnaces burning. Selling them the raw power,
• 3:46
  essentially. Yes. And they're selling them a
• 3:48
  billion dollars worth of coal every single month.
• 3:51
  But my question to you is like, how on earth
• 3:53
  is that sustainable? I mean, is a $900 billion
• 3:56
  valuation just a bubble when your overhead is
• 3:59
  a billion dollars a month? Well, what's fascinating
• 4:01
  here is what that dynamic actually reveals about
• 4:05
  the mechanics of the AI economy itself. It is
• 4:09
  entirely unprecedented. When you ask if it's
• 4:11
  sustainable, you really have to understand why
• 4:13
  those costs are so high in the first place. Right,
• 4:15
  because it's not like normal software. No, this
• 4:17
  isn't like traditional software at all. How so?
• 4:20
  Because, I mean, if I build a traditional sauce
• 4:22
  application, my cloud hosting bill doesn't look
• 4:25
  anything like that. even if I have millions of
• 4:27
  users? Because traditional software is essentially
• 4:31
  deterministic. You write the code, you compile
• 4:34
  it, host it on a server somewhere, and when a
• 4:37
  user clicks a button, the server just executes
• 4:40
  a pre -written set of instructions. So the computational
• 4:43
  cost to run it for one user versus a million
• 4:45
  users, it scales incredibly efficiently. Right.
• 4:48
  It's just retrieving things. Exactly. Generative
• 4:51
  AI fundamentally breaks that scaling model. So
• 4:54
  every time someone queries an AI, it's like a
• 4:56
  much heavier lift for the system. A massively
• 4:58
  heavier lift. It's a process called inference.
• 5:01
  So every single time you ask a large language
• 5:03
  model a question, it isn't just retrieving a
• 5:07
  static file. It's actually thinking. Well, calculating.
• 5:10
  Right. Those server farms have to run these incredibly
• 5:12
  complex probabilistic calculations across hundreds
• 5:16
  of billions of parameters in real time just to
• 5:19
  predict the next word and then the next word
• 5:21
  and the next. And that requires an immense amount
• 5:23
  of continuous electricity and highly specialized
• 5:26
  processors, you know, the GPUs. You are essentially
• 5:29
  renting a supercomputer for every single individual
• 5:32
  query. Which totally explains the billion dollar
• 5:35
  SpaceX bill. Yeah, the compute power required
• 5:38
  just to keep the lights on and answer user prompts
• 5:41
  is astronomical. And that's not even talking
• 5:44
  about the compute required to train the next
• 5:46
  generation of models. Right. That's a whole other
• 5:48
  massive expense. Exactly. So to answer your question
• 5:50
  about sustainability, a $900 billion valuation
• 5:54
  means the market truly believe these models will
• 5:57
  restructure the global economy so thoroughly
• 6:00
  that a billion dollar monthly overhead. will
• 6:03
  eventually just look like a bargain. Like a drop
• 6:05
  in the bucket. Right. But in the short term,
• 6:07
  those staggering bills force these companies
• 6:10
  to operate at a breakneck, just relentless speed.
• 6:13
  They can't afford to slow down. They really can't.
• 6:15
  They have to keep growing, keep deploying, and
• 6:18
  keep monetizing at maximum velocity just to outrun
• 6:21
  their own infrastructure costs. Which naturally
• 6:23
  leads us to the next major point in the source
• 6:25
  material. And here's where it gets really interesting,
• 6:28
  because... If these companies are building the
• 6:30
  future at this blistering speed due to financial
• 6:33
  pressure, you have to ask, you know, who is directing
• 6:36
  traffic? Who's in charge, yeah. Right. Who is
• 6:38
  setting the rules of the road? And according
• 6:40
  to the Kinsoff report, the answer seems to be
• 6:43
  that policy is entirely unsettled. The policy
• 6:46
  vacuum is, I'd argue, a defining characteristic
• 6:49
  of this current era. Yeah. The report highlights
• 6:52
  this very specific event. President Trump abruptly
• 6:55
  canceled the signing of a major AI executive
• 6:58
  order. Right. Just pull the plug on it. And the
• 7:00
  explicit reason provided for canceling it was
• 7:02
  that it risked undermining America's competitive
• 7:05
  edge. And just to be clear, looking at this strictly
• 7:08
  as a classic technological dilemma, completely
• 7:10
  setting aside politics, this feels like the ultimate
• 7:14
  double -edged sword. Oh, absolutely. Because
• 7:16
  on one hand, you obviously want to win the global
• 7:18
  economic race. But on the other hand, by intentionally
• 7:21
  removing the speed limits, aren't we kind of
• 7:24
  driving blind into this incredibly powerful technology?
• 7:27
  Well, if we connect this to the bigger picture,
• 7:29
  what you are observing is essentially a classic
• 7:33
  game theory scenario. But it's playing out on
• 7:36
  a geopolitical scale. Right. It is that fundamental
• 7:39
  historical tension between regulation and innovation.
• 7:42
  But now it's amplified by the sheer speed of
• 7:46
  AI development. Nations are acutely aware that
• 7:50
  AI is not just another software product. It's
• 7:53
  way bigger than that. It is viewed as the foundational
• 7:55
  technology of the next century. much like the
• 7:58
  electrical grid or the Internet itself. I understand
• 8:00
  the desire for dominance, obviously, but surely
• 8:03
  establishing some basic federal -level guardrails
• 8:06
  would stabilize the market. Like, doesn't a total
• 8:09
  lack of rules invite disaster for the companies
• 8:11
  who are actually trying to build in this space?
• 8:13
  I mean, from a purely domestic standpoint, yes.
• 8:16
  Clear regulations provide market certainty, which
• 8:18
  businesses love. Right. But the issue here is
• 8:21
  international parity. Okay. If you heavily regulate
• 8:23
  your domestic AI industry to ensure it is perfectly
• 8:26
  safe, thoroughly tested, ethically constrained,
• 8:29
  you inherently slow down the development cycle.
• 8:32
  Because all that testing takes a lot of time.
• 8:33
  Exactly. And in a vacuum, that's the responsible
• 8:36
  thing to do. But we don't live in a vacuum. The
• 8:40
  calculation being made is that if you slow down
• 8:42
  your own tech giants, international competitors
• 8:45
  who, frankly, do not share your regulatory concerns
• 8:48
  will just race ahead. And they will capture all
• 8:51
  the economic benefits, the intellectual property,
• 8:53
  and, I mean, the military applications. Precisely.
• 8:57
  The state apparatus is essentially looking at
• 8:59
  the landscape and deciding that the risk of falling
• 9:02
  behind is fundamentally greater than the risk
• 9:05
  of moving too fast. That is exactly why the executive
• 9:08
  order was canceled, to maintain that competitive
• 9:10
  edge. So the immediate takeaway for you listening
• 9:12
  to this deep dive is that you cannot wait for
• 9:15
  the government to figure out the rules of AI.
• 9:17
  You're on your own. Pretty much. You have to
• 9:20
  navigate this landscape under the assumption
• 9:22
  that federal policy will remain entirely unsettled
• 9:25
  for the foreseeable future. Which means the responsibility
• 9:27
  for safety and security, it doesn't lie with
• 9:30
  some regulatory body. It falls entirely on the
• 9:33
  companies themselves. Like, we're relying on
• 9:35
  the AI models to just police themselves. And
• 9:37
  that transition of responsibility, that leads
• 9:40
  us directly into the most concerning technical
• 9:43
  revelations in the Kinsoft report. Yeah, because
• 9:46
  if the government is hitting pause on regulations,
• 9:49
  we are placing this immense amount of trust in
• 9:52
  the internal defenses of these commercial tools.
• 9:55
  But the report reveals that those internal defenses
• 9:57
  are, I mean, shockingly fragile. They really
• 10:01
  are. It says researchers reported that they could
• 10:03
  remove the safety guardrails from major AI models
• 10:06
  in a matter of minutes. Just minutes. Yeah, it's
• 10:09
  wide. And once those guardrails were removed,
• 10:11
  the models just happily answered prompts they
• 10:13
  were explicitly supposed to refuse. It completely
• 10:16
  shatters this illusion of inherent model safety
• 10:18
  that is so often marketed to the public. Honestly,
• 10:22
  I really have to push back on the assurances
• 10:24
  we constantly hear from the tech industry. They
• 10:26
  tell enterprise customers that these models are
• 10:28
  aligned, that they have these strict safeguards,
• 10:31
  that they won't generate malicious code or expose
• 10:33
  sensitive corporate data. Right. That's the sales
• 10:36
  pitch. But if it only takes a team of researchers
• 10:38
  a few minutes to break those rules, are these
• 10:41
  guardrails essentially just security theater?
• 10:44
  To me, it sounds like putting a flimsy piece
• 10:47
  of yellow caution tape in front of a massive
• 10:49
  bank vault. You can issue a press release saying
• 10:51
  the vault is protected, but like anyone can just
• 10:54
  duck under the tape. This raises an important
• 10:56
  question, though, which is what exactly is a
• 10:59
  guardrail in the context of a large language
• 11:03
  model? Good question. I think the general public
• 11:05
  and honestly, even many IT professionals, they
• 11:08
  imagine a physical barrier or like a hard coded
• 11:11
  firewall. Right. Like an if then statement. Exactly.
• 11:14
  They imagine a line of code that literally says
• 11:17
  if user asks for X, block the request. Yeah.
• 11:21
  But that is simply not how. neural networks operate.
• 11:24
  How do they operate then? Because clearly the
• 11:25
  word guardrail is doing a lot of heavy lifting
• 11:27
  here. Yeah, it really is. A guardrail in an LLM
• 11:30
  is essentially just another layer of training
• 11:32
  data. Wait, really? Just data? Yeah. It is a
• 11:35
  set of statistical weights that biases the model
• 11:39
  toward generating a refusal like saying, you
• 11:42
  know, I cannot help you with that when it recognizes
• 11:45
  certain patterns of words. Oh, I see. It is soft
• 11:48
  instruction layered over a probabilistic engine.
• 11:51
  And the fundamental problem with software instruction
• 11:54
  that operates entirely on human language is that
• 11:56
  human language is infinitely flexible. Right,
• 11:59
  of course, because there are a million different
• 12:00
  ways to ask the exact same question. Exactly.
• 12:03
  So clever researchers or malicious actors, they
• 12:07
  use what are called adversarial prompts or jailbreaks.
• 12:11
  Right, we've heard of jailbreaks. Yeah, and they
• 12:12
  don't hack the code, right? They logically trap
• 12:15
  the model. They use complex scenarios or role
• 12:18
  -playing prompts or weird hypotheticals to confuse
• 12:20
  the model's pattern recognition. So they trick
• 12:23
  it. Basically. Because the model is just predicting
• 12:25
  the next most likely token based on its training.
• 12:27
  If you can structure a prompt that makes a dangerous
• 12:30
  response statistically more likely than a refusal
• 12:32
  response, the guardrail just dissolves. You essentially
• 12:35
  just talk the math out of being safe. You bypass
• 12:38
  the statistical bias exactly, which is why there
• 12:41
  is a massive foundational difference between
• 12:43
  a model's built -in safety rules and actual corporate
• 12:46
  data security. Okay, how so? Well, the model's
• 12:49
  internal safety is designed to stop it from generating
• 12:51
  generally harmful content, like bad instructions
• 12:54
  or toxic speech. It is not a bespoke, rigid security
• 12:58
  architecture designed to protect your company's
• 13:01
  proprietary data from a targeted extraction attempt.
• 13:04
  Wow. Which brings us to the very... explicit
• 13:07
  advice Kinsoft gives to businesses in this report.
• 13:09
  They state, if you let staff use AI tools, assume
• 13:14
  guardrails can be bypassed. Put your own policies
• 13:17
  and monitoring around how tools are used with
• 13:19
  company data. And that is the crucial pivot every
• 13:22
  single organization needs to make in their threat
• 13:24
  modeling right now. You have to assume it's vulnerable.
• 13:26
  Yes. You cannot trust the AI's internal safety
• 13:29
  features to protect your environment. If you
• 13:31
  are integrating these tools into your daily workflows,
• 13:34
  you have to operate under the baseline assumption
• 13:36
  that the AI will at some point behave exactly
• 13:39
  as a malicious actor or an accidental insider
• 13:41
  wants it to. That's terrifying, honestly. It
• 13:44
  is, which is why you must rely on your own internal
• 13:47
  monitoring, your own data loss prevention tools
• 13:50
  and strict traditional access controls. OK, so
• 13:53
  just to summarize the logic here, because the
• 13:56
  AI's internal defenses are easily bypassed through
• 13:59
  language manipulation, we have to fall back on
• 14:02
  our traditional cybersecurity defenses. We have
• 14:05
  to basically trust the enterprise security tools
• 14:08
  we've relied on for the past decade to catch
• 14:10
  any malicious behavior. Yes, that is the standard.
• 14:13
  But as the report details in its final section,
• 14:16
  that foundational layer of legacy cybersecurity,
• 14:20
  it's cracking under pressure as well. And this
• 14:23
  is the part of the report that completely changes
• 14:25
  the threat model for every IT department. It's
• 14:27
  massive. Because the vulnerability isn't just
• 14:29
  limited to brand new experimental AI tools where
• 14:33
  we kind of expect some rough edges. We are dealing
• 14:37
  with the reality that our established cybersecurity
• 14:40
  infrastructure is being deeply compromised. The
• 14:43
  report highlights that a major trusted security
• 14:45
  vendor, Trellix, was breached, but they didn't
• 14:48
  just lose customer data, which is bad enough.
• 14:51
  Attackers gained access to the source code behind
• 14:54
  Trellix's own security products. And in the cybersecurity
• 14:57
  world, unauthorized access to a vendor's source
• 15:00
  code is a worst case scenario. It is a fundamental
• 15:03
  breach of trust. So what does this all mean?
• 15:06
  Let me offer an analogy to make this kind of
• 15:08
  abstract threat concrete for anyone listening.
• 15:10
  Go for it. Imagine you own a highly secure facility,
• 15:13
  right? And you hire a premier security company
• 15:16
  to install these unpickable, state -of -the -art
• 15:18
  locks on all your doors. You feel perfectly safe.
• 15:21
  Naturally. But then you find out that the company
• 15:23
  manufacturing those physical locks just had their
• 15:26
  master blueprint stolen by a highly organized
• 15:28
  syndicate of burglars. Oh, man. The burglars
• 15:31
  aren't breaking into your building right now,
• 15:32
  but they're sitting in a warehouse somewhere.
• 15:34
  studying the exact schematics of your lock, figuring
• 15:38
  out precisely where the gears catch and where
• 15:40
  the microscopic weaknesses are. That analogy
• 15:42
  captures the severity perfectly. Because when
• 15:45
  attackers acquire the source code of a security
• 15:47
  tool, they aren't just looking for a quick financial
• 15:49
  gain like a standard ransomware attack. They're
• 15:52
  playing the long game. Exactly. They are stealing
• 15:55
  the rulebook. They can compile that proprietary
• 15:57
  code in their own isolated environments, reverse
• 16:00
  engineer the detection algorithms, and methodically
• 16:03
  hunt for zero -day vulnerabilities. Which are
• 16:06
  the flaws nobody knows about yet. Right. Weaknesses
• 16:08
  that the vendor doesn't even know exist yet.
• 16:10
  But doesn't that render the entire concept of
• 16:13
  legacy cybersecurity obsolete? Like, if the attackers
• 16:16
  have the blueprints to the defense, why even
• 16:19
  bother paying for the software? It doesn't render
• 16:21
  it obsolete. Because you still need defenses
• 16:23
  against the millions of automated known attacks
• 16:26
  that happen every single day. Okay, fair point.
• 16:28
  But it means legacy cybersecurity is no longer
• 16:31
  a silver bullet against advanced threats. They
• 16:34
  can analyze exactly how the security software
• 16:36
  looks for malware and then design their own malicious
• 16:39
  code to look like something else entirely. So
• 16:41
  they know exactly how to disguise themselves.
• 16:44
  Yes. It is weaponizing. the defense mechanisms
• 16:47
  against the defenders. It's the terrifying concept
• 16:50
  of supply chain and vendor vulnerability. Because
• 16:53
  in today's digital economy, you aren't just responsible
• 16:55
  for your own perimeter anymore. No, absolutely
• 16:58
  not. You could have flawless internal policies.
• 17:01
  Your employees could be perfectly trained to,
• 17:03
  you know, never click a phishing link. But if
• 17:07
  the software you purchase to protect your endpoints
• 17:09
  is fundamentally flawed because its blueprints
• 17:11
  were stolen, your entire network is exposed.
• 17:15
  And the Kinsoft report doesn't just deal in hypotheticals
• 17:18
  here either. They provide these stark examples
• 17:20
  of the collateral damage that occurs when these
• 17:22
  interconnected systems fail. Real world consequences.
• 17:25
  Exactly. We aren't just talking about software
• 17:28
  companies losing data. They highlighted West
• 17:31
  Pharmaceutical Services, a major manufacturer
• 17:33
  of drug delivery components. They spent two full
• 17:37
  weeks recovering from a cyber attack that completely
• 17:39
  crippled its systems. Two weeks. Two weeks of
• 17:42
  downtime for a pharmaceutical manufacturer. ripples
• 17:44
  out. It impacts the global supply chains. It
• 17:47
  delays shipments to hospitals. It ultimately
• 17:49
  impacts patient care. It perfectly demonstrates
• 17:52
  how digital vulnerabilities can just halt physical
• 17:55
  infrastructure. And the report also notes that
• 17:57
  cyber attacks on the education sector jumped
• 17:59
  more than 50 percent month on month. So we are
• 18:02
  talking about schools, universities and research
• 18:05
  institutions being actively targeted now. It
• 18:08
  just proves that absolutely no industry is insulated
• 18:10
  from this. No industry is insulated because the
• 18:13
  underlying digital foundation is a shared resource.
• 18:17
  Whether you are a critical pharmaceutical giant,
• 18:20
  a local public school district, or a $900 billion
• 18:23
  AI startup, you are all relying on a very small,
• 18:27
  consolidated handful of security vendors, cloud
• 18:30
  providers, and software supply chains. We're
• 18:33
  all in the same boat. Exactly. When one of those
• 18:35
  foundational pillars is compromised, like a vendor
• 18:38
  source code being breached, the shock waves hit
• 18:40
  everyone simultaneously. Which brings us to the
• 18:43
  vital question every single listener needs to
• 18:45
  take back to their team tomorrow morning. What
• 18:47
  happens if one of our trusted security vendors
• 18:49
  gets breached? It's a complete paradigm shift,
• 18:52
  really, because we spend all our time running
• 18:53
  these tabletop exercises asking, what happens
• 18:56
  if we get breached? Right. You focus on your
• 18:58
  own walls. But you have to war game this scenario
• 19:01
  where the tools you implicitly trust are turned
• 19:04
  against you or just fail silently. Wow. Do you
• 19:08
  have redundant detection systems? Can you isolate
• 19:11
  sensitive network segments if your primary endpoint
• 19:14
  protection is compromised by a supply chain attack?
• 19:17
  If the answer is no, you have a massive critical
• 19:20
  blind spot in your architecture. Which makes
• 19:22
  Kinsoff's final piece of advice in the report
• 19:24
  incredibly poignant. They simply state, stay
• 19:28
  patched, stay skeptical. It is the only rational
• 19:31
  operational posture to take in a landscape this
• 19:35
  volatile. You must relentlessly patch your systems
• 19:38
  because you have to close the known vulnerabilities.
• 19:40
  But you must remain deeply skeptical of the software
• 19:43
  you deploy because you know the unknown vulnerabilities
• 19:45
  are currently being mapped by highly capable
• 19:47
  adversaries. It has been a truly wild journey
• 19:50
  on this deep dive. Let's just briefly recap the
• 19:52
  sheer scale of the terrain we've covered today.
• 19:54
  It's a lot of ground. It really is. We started
• 19:57
  with the absolute economic euphoria of anthropics.
• 20:01
  Nine hundred billion dollar valuation balanced
• 20:05
  against the staggering reality of SpaceX's billion
• 20:08
  dollar a month compute bills just to keep the
• 20:11
  inference engines running. Right. The coal for
• 20:13
  the factories. Exactly. Then we looked at a geopolitical
• 20:16
  landscape where policy is intentionally left
• 20:18
  unsettled to win a global race, effectively removing
• 20:21
  the speed limits on development. We discovered
• 20:24
  that the internal state. statistical guardrails
• 20:26
  of these AI models can be bypassed by researchers
• 20:29
  using adversarial language in mere minutes. Just
• 20:32
  talking the math out of being safe. Right. And
• 20:34
  finally, we saw that our traditional defenders,
• 20:36
  security vendors like Trellix, are having their
• 20:39
  own source code breached, leading to devastating
• 20:41
  real -world consequences for critical industries
• 20:44
  like pharmaceuticals and education. It is a profound
• 20:47
  and... honestly, highly volatile convergence
• 20:50
  of technological trends. And this is exactly
• 20:53
  why you should care. Whether you are prepping
• 20:56
  for a high -stakes strategy meeting, trying to
• 20:58
  protect your own small business network, or you're
• 21:00
  just incredibly curious about the mechanics of
• 21:03
  the future, you have to operate with this ghoul
• 21:06
  reality in mind. You really do. You cannot just
• 21:09
  blindly trust the glossy, optimistic brochures
• 21:13
  of the AI revolution, and you cannot assume your
• 21:16
  legacy security tool. are impenetrable fortresses,
• 21:19
  you have to stay patched and you have to stay
• 21:21
  deeply skeptical of the infrastructure you use
• 21:24
  every single day. I want to leave you with a
• 21:26
  final thought to mull over, something that connects
• 21:28
  the two very distinct halves of our conversation
• 21:31
  today. Okay, let's hear it. We established that
• 21:33
  clever researchers can bypass AI safety guardrails
• 21:36
  using language manipulation in a matter of minutes,
• 21:39
  right? Right. We also established that hackers
• 21:41
  now have the source code to hunt for unseen weaknesses
• 21:44
  in major security vendors like Trellix. So what
• 21:47
  happens when those two realities merge? Oh, wow.
• 21:51
  What happens the day a brilliant guardrail -free
• 21:54
  AI model is deliberately prompted by a malicious
• 21:57
  actor to autonomously hunt through that stolen
• 22:02
  security source code, when an AI that can process
• 22:05
  millions of lines of code a minute is tasked
• 22:08
  with finding zero -day vulnerabilities that human
• 22:10
  hackers would take years to spot. Oh, man. It's
• 22:13
  like handing a supercomputer to the burglar who
• 22:15
  already stole the blueprints to your locks. The
• 22:18
  pace of discovery is going to accelerate faster
• 22:20
  than we can even comprehend. We are building
• 22:22
  the most incredible technological engine in human
• 22:25
  history, but we desperately need to start paying
• 22:27
  a lot more attention to the fragile foundation
• 22:28
  we are laying down underneath it. Thank you so
• 22:30
  much for joining us on this deep dive. We'll
• 22:32
  catch you on the next one.