Latest / Cybersecurity Under Pressure. Real Attacks, Real Lessons / Kinetic AI and Critical Infrastructure: Connecting Cyber and Physical Risk

Transcript

Kinetic AI and Critical Infrastructure: Connecting Cyber and Physical Risk · Cybersecurity Under Pressure. Real Attacks, Real Lessons

• 0:47
  So the guy looks at the logs right?
• 0:49
  And I kid you not, the default password for this entire water
• 0:53
  treatment facilities remote access was.
• 0:56
  I mean it was literally just the word admin spelled.
• 0:58
  Backwards. Oh no, Nimda.
• 1:00
  Nimda, yes. Which is it's so incredibly
• 1:03
  ironic because Nimda was that massive, globally destructive
• 1:07
  computer worm back in, what, the early 2000s?
• 1:10
  It's like they were actively trying to tempt fate.
• 1:12
  Right. Or just like taunt the universe?
• 1:14
  Yeah, pretty much. It's the kind of thing you'd
• 1:16
  absolutely laugh at if it wasn't connected to, you know, a
• 1:19
  municipal water supply that thousands of actual human beings
• 1:21
  drink from everyday. Just unbelievable.
• 1:23
  But anyway, you know what I've been thinking about this week?
• 1:27
  The story of an incident that started with a technical alert
• 1:30
  and ended with a lesson that every CI SO should be asking
• 1:33
  about. Oh yeah, it is definitely a
• 1:35
  story that forces a complete ground up recalibration of how
• 1:39
  we think about digital defense because the old models, I mean
• 1:42
  they simply do not apply anymore.
• 1:44
  We are in completely uncharted territory.
• 1:46
  Exactly. And that's our mission for this
• 1:48
  deep dive. We are going to explore the
• 1:51
  terrifying collision of operational technology,
• 1:54
  artificial intelligence and physical infrastructure.
• 1:57
  The physical aspect is what really changes the game.
• 1:59
  Here it changes everything. We're going to be unpacking the
• 2:02
  latest directives, specifically looking at the April 2026 CISA
• 2:08
  guidance, and we're putting that alongside some incredibly dense
• 2:12
  but honestly fascinating intelligence reports from Suzu
• 2:15
  Labs in Shieldworks. Plus the historical context.
• 2:18
  Right. Plus, we're looking at some
• 2:19
  truly sobering historical postmortems from NIST and
• 2:22
  Anissa. So for everyone listening, I
• 2:24
  want you to imagine you are sitting in the control room of a
• 2:26
  major power plant. We're looking at what happens
• 2:30
  when the digital world violently impacts your physical reality.
• 2:33
  And we really, really need to emphasize that word violent,
• 2:36
  because when we talk about operational technology or OT, we
• 2:39
  aren't talking about losing access to your e-mail for a few
• 2:42
  hours. Or like a customer database
• 2:45
  getting encrypted and you have to pay ransom.
• 2:47
  Exactly. This isn't data loss.
• 2:48
  We're talking about the kinetic world.
• 2:50
  We are talking about massive blast furnaces, robotic
• 2:54
  manufacturing arms swinging around National Rail networks,
• 2:58
  petrochemical refineries. The stakes are fundamentally
• 3:02
  different because the consequences they aren't
• 3:04
  measured in lost revenue, they're measured in
• 3:06
  environmental disasters and, quite frankly, human lives.
• 3:09
  Yeah, which is a very heavy place to start, but we have to
• 3:12
  set the table. Yeah.
• 3:13
  So let's start at the foundation of how these places used to be
• 3:16
  protected, which ironically is a foundation that essentially no
• 3:20
  longer exists. The air gap.
• 3:21
  The air gap. For decades, the entire defense
• 3:24
  strategy for these massive physical systems was just this
• 3:27
  one concept. And the logic was incredibly
• 3:30
  straightforward, right? You just you don't plug the
• 3:32
  nuclear plant or the water treatment facility into the
• 3:34
  public Internet. Problem solved.
• 3:36
  That was the premise, yes. The core assumption for the
• 3:38
  longest time was that your IT network.
• 3:40
  So your corporate emails, your billing systems, the HR portals,
• 3:44
  that entire world was completely both physically and logically
• 3:48
  severed from your OT network. There were two different
• 3:50
  planets. Completely different planets.
• 3:52
  The OT network is where you find the control loops, the
• 3:55
  supervisory control and data acquisition stuff, the scatter
• 3:59
  platforms and the programmable logic controllers, the PLC's.
• 4:04
  And for anyone who isn't deep in the engineering weeds, those
• 4:07
  PLC's are basically just specialized, heavily ruggedized
• 4:10
  computers, right? Yeah, exactly.
• 4:12
  They are the brains on the factory floor that actually send
• 4:14
  the electrical signals to open a physical valve or spin up a
• 4:17
  turbine and the air gap. There was literally no wire
• 4:21
  connecting those sensitive PLCS to the corporate Wi-Fi or the
• 4:25
  outside world. Right.
• 4:26
  But if you walk onto a factory floor today, that air gap is a
• 4:30
  total myth. It's completely collapsed.
• 4:32
  And the crazy part is it wasn't hackers who destroyed it, It was
• 4:35
  the executive. Oh absolutely.
• 4:37
  It was the C-Suite pushing for what they call Industry 4 Point
• 4:40
  O. Which sounds great in the
• 4:41
  boredom. It sounds fantastic in a
• 4:44
  PowerPoint presentation. It was this relentless drive for
• 4:47
  efficiency and data-driven decision making.
• 4:50
  I mean, imagine you are a plant manager today.
• 4:53
  You have executives constantly demanding real time predictive
• 4:57
  maintenance, remote monitoring capabilities, automated supply
• 5:01
  chain integration. They want all the data all the
• 5:04
  time. Exactly.
• 5:05
  And if you want a cloud based artificial intelligence to
• 5:08
  analyze vibration data and predict exactly when a specific
• 5:12
  ball bearing inside a multi $1,000,000 turbine is going to
• 5:16
  fail, well, the acoustic sensor on that bearing must eventually
• 5:20
  send its data out to the cloud. It has to have a path out.
• 5:23
  Right. You cannot have localized
• 5:25
  analytics anymore. The business demanded
• 5:27
  connectivity. So the engineers bridged the air
• 5:29
  gap. They ran the cables.
• 5:31
  Yeah, I mean, my toaster apparently needs a direct Wi-Fi
• 5:33
  connection to brown bread properly now.
• 5:35
  So of course the critical cooling pumps at a chemical
• 5:38
  plant need to be accessible from some vendors smartphone in
• 5:41
  another country. It's the exact same underlying
• 5:43
  logic, just with catastrophic consequences.
• 5:45
  It's just we willingly traded existential security for like
• 5:50
  dashboard convenience. It is a pretty bleak way to look
• 5:53
  at it, but operationally that is exactly what happened.
• 5:55
  We basically punched a million tiny holes in the fortress wall
• 5:59
  just to let the data out without ever really realizing what could
• 6:03
  come back in through those same holes.
• 6:05
  Which brings us directly to the government's response to this
• 6:08
  mess Yeah the April 2020 sixth see by Bodie 26 O 4 directive
• 6:14
  yes and from reading the sources, this isn't just another
• 6:16
  boring government memo. This is a monumental shift in
• 6:20
  federal cybersecurity policy. It is huge.
• 6:23
  CISA basically looked at this highly connected post air gap
• 6:27
  landscape and realized that our entire framework for
• 6:30
  prioritizing vulnerabilities was just fundamentally broken.
• 6:33
  Because of the CVSS scores, right?
• 6:35
  It's the common vulnerability scoring system.
• 6:37
  Exactly. If you're a security engineer,
• 6:39
  you live and die by these numbers.
• 6:40
  A 10 out of 10 means you drop whatever you're doing, you
• 6:43
  cancel your weekend, and you fix it.
• 6:44
  Under the old context, that's exactly how it worked.
• 6:47
  A severity rating of, say, a CVSS 9.8 was treated as a Heron
• 6:52
  fire emergency across the board, completely regardless of the
• 6:55
  actual context of the device. Which makes no sense when you
• 6:58
  think about it physically. Right.
• 6:59
  Think about the physical reality of a network.
• 7:02
  A 9.8 vulnerability on an isolated internal database
• 7:06
  that's sitting in a locked basement on a segmented VLAN
• 7:09
  that received the exact same urgency as a 9.8 vulnerability
• 7:13
  on AVPN gateway that has a direct Internet connection and
• 7:18
  is actively being scanned by Fed actors in the wild.
• 7:21
  Which means federal agencies who, let's be honest, always
• 7:24
  have limited time, limited budgets, they're understaffed.
• 7:28
  They were running around patching the wrong things first.
• 7:30
  Constantly. They were treating a theoretical
• 7:32
  risk in a locked vault the same as like a a burglar actively
• 7:37
  kicking in the front door. That categorization mismatch was
• 7:41
  a massive operational vulnerability in and of itself.
• 7:44
  I mean, defenders were completely exhausted from alert
• 7:46
  fatigue. They were patching low risk
• 7:48
  internal assets while their Internet exposed critical
• 7:51
  infrastructure just sat there wide open.
• 7:53
  But the SISA directive explicitly changes that dynamic,
• 7:56
  right? Right.
• 7:56
  It forces a three day critical patch mandate specifically for
• 8:00
  those Internet exposed assets facing AI driven threats.
• 8:03
  It does. But wait, I have to push back
• 8:05
  here because reading that made me crazy.
• 8:08
  How can any organization patch perfectly in three days?
• 8:12
  I mean, think about it from the perspective of an operator on
• 8:15
  the floor. It's a huge.
• 8:17
  Ass. In a normal IT environment,
• 8:19
  forcing a patch in 72 hours is hard enough, but in an OT
• 8:23
  environment, patching A programmable logic controller
• 8:27
  might mean shutting down a continuous manufacturing line,
• 8:29
  yes. It absolutely might.
• 8:31
  It requires scheduled downtime, safety reviews, vendor approvals
• 8:35
  to make sure the patch doesn't break the warranty.
• 8:38
  Three days isn't just aggressive in the OT world, it sounds
• 8:41
  clinically insane. From a legacy operational
• 8:44
  standpoint, you are completely correct.
• 8:45
  It sounds impossible. The engineers are probably
• 8:48
  tearing their hair out over it, but it is a necessary, almost
• 8:51
  desperate reaction to the reality outlined in the Suzu
• 8:54
  Labs intelligence reporting. OK, let's get into the Suzu lab
• 8:57
  stuff because this blew my mind. The Suzu Labs analyst point out
• 9:01
  a terrifying metric they found that the meantime to exploit has
• 9:04
  actually gone negative. OK, stop.
• 9:06
  You have to explain the mechanics of that, because
• 9:08
  logically, how can a timeline be negative?
• 9:11
  You can't exploit something before it exists.
• 9:13
  It sounds like time travel, but it comes down to how artificial
• 9:17
  intelligence has fundamentally weaponized vulnerability
• 9:20
  discovery. Let's look at the old model for
• 9:22
  a second. The way things used to work.
• 9:24
  Right, a software vendor discovers a flaw.
• 9:28
  They release a patch on, say, a Tuesday, and then human
• 9:32
  attackers would take weeks or sometimes even months to reverse
• 9:36
  engineer that patch, figure out what changed, and manually write
• 9:39
  an exploit to take advantage of the systems that hadn't applied
• 9:43
  the patch yet. Right, so the defenders had a
• 9:44
  head start, the clock started ticking, but they had a buffer.
• 9:47
  Exactly. But now adversaries are feeding
• 9:50
  those patch binaries directly into AI models designed for
• 9:53
  automated binary diffing. Automated diffing OK.
• 9:55
  The AI compares the old vulnerable file with the new
• 9:59
  patched file. It instantly maps out the
• 10:01
  assembly code that changed. Perhaps you know a new memory
• 10:04
  bounce check was added by the developer, so the.
• 10:06
  AI immediately sees where the Band-Aid was placed.
• 10:08
  Exactly. The AI instantly understands
• 10:11
  that the old version has a buffer overflow vulnerability at
• 10:14
  that specific memory address. It then uses a large language
• 10:19
  model to trace the execution path back to a user input field,
• 10:22
  and literally within hours it automatically generates a Python
• 10:25
  script to send a maliciously crafted payload to execute shell
• 10:29
  code. Wow.
• 10:30
  They are weaponizing the flaw before the defenders even have
• 10:33
  time to download the patch, let alone test it in a sandbox
• 10:36
  environment to see if it breaks the factory line.
• 10:39
  The attackers are moving faster than linear time allows.
• 10:41
  The human security teams change Advisory Board to react so.
• 10:44
  They are getting hit before the meeting about the patch even
• 10:48
  happens. Exactly.
• 10:49
  Hence negative time to exploit that.
• 10:51
  Is deeply, deeply unsettling, and the proof isn't theoretical
• 10:54
  here. It's already in the wild, like
• 10:56
  the March 2020 Sixth Strucker cyberattack.
• 10:59
  Yes, the. Medical technology giant a
• 11:00
  cyberattack wipe devices across 79 countries simultaneously 70.
• 11:04
  Nine countries at once that. Is the true scale and speed of
• 11:09
  AI driven exploitation. It wasn't a slow creeping
• 11:12
  infection where you see one server go down in London and
• 11:15
  then another one goes down a day later in Tokyo.
• 11:18
  It was a global synchronized wipe executed the exact moment
• 11:22
  the vulnerability was mapped by the AI.
• 11:24
  So if. You're a C ISO and the exploit
• 11:26
  window is measured in hours. The CISA says you have three
• 11:29
  days to patch. You're technically already
• 11:32
  behind before the clock even starts.
• 11:33
  You are. But here's where it gets
• 11:35
  incredibly nuanced with this mandate, and I think a lot of
• 11:38
  people miss this. The CISA directive doesn't just
• 11:40
  blindly say, deploy the patch in three days.
• 11:44
  It mandates a three day window to conduct mandatory forensic
• 11:48
  triage before applying the patch.
• 11:50
  This. Is the most crucial and often
• 11:52
  entirely misunderstood part of the directive.
• 11:55
  It speaks to a very specific, documented operational risk.
• 11:58
  If you have an Internet exposed system that has been vulnerable
• 12:02
  for even a few hours in this new AI paradigm, you must assume it
• 12:06
  has already been compromised. You have to.
• 12:08
  Operate into the assumption that prevention has already failed,
• 12:10
  which is a really bitter pill for traditional network
• 12:14
  defenders to swallow, right? They want to believe their
• 12:16
  firewalls are working. But look at the mechanics of an
• 12:20
  attack. If you just blindly apply the
• 12:22
  patch on day 2, what are you actually doing?
• 12:25
  From a technical standpoint, you're.
• 12:27
  Just updating a file you. Are overwriting the vulnerable
• 12:30
  executable with a secure one and usually rebooting the system.
• 12:34
  But you aren't checking if the adversary is already inside the
• 12:38
  operating system. If the system is already back
• 12:40
  doored, patching without investigation simply allows the
• 12:44
  adversary to maintain their persistence because.
• 12:47
  They've already moved laterally. Exactly during that brief window
• 12:51
  of vulnerability, the automated exploit might have dropped a
• 12:54
  rootkit deep into the kernel, or created new unauthorized
• 12:58
  administrative credentials in Active Directory, or even
• 13:01
  modified the firmware of a network card and a.
• 13:03
  Patch doesn't look for any of that.
• 13:05
  None. Of those things are fixed by
• 13:06
  patching A vulnerable web service.
• 13:08
  The malicious implants will easily survive the patch cycle
• 13:12
  and the subsequent reboot I was trying.
• 13:13
  To think of a way to visualize this while I was reading the
• 13:16
  reports, it's like it's like discovery.
• 13:18
  Your front door lock is broken, right?
• 13:20
  If you just blindly replace the lock on the front door, you've
• 13:24
  secured the perimeter, but you haven't checked to see if the
• 13:26
  killer is already upstairs in the guest bedroom.
• 13:28
  That is, you actually. Just lock yourself inside with
• 13:31
  them that. Is a very visceral analogy, but
• 13:33
  mechanically it is spot on. The Sisa mandate forces
• 13:37
  defenders to shift the fundamental question.
• 13:40
  Instead of asking how quickly can we push this patch, security
• 13:44
  teams must now ask can we prove this system wasn't compromised
• 13:49
  during the exposure window you have?
• 13:51
  To prove the negative you have. To hunt continuously through the
• 13:54
  logs and the memory space, you have to compress the dwell time
• 13:58
  of the attacker. As the Suzu Labs report
• 14:01
  concludes, the organizations that survive this new era won't
• 14:04
  be the ones that patch perfectly.
• 14:06
  Because you can't. Nobody.
• 14:08
  Patches perfectly at -7 day speed.
• 14:11
  The survivors will be the ones who assume exposure, conduct the
• 14:14
  forensic triage, and use their own defensive AI to scale their
• 14:18
  hunting capabilities inside the network.
• 14:20
  But. Wait, this brings up a massive
• 14:22
  glaring logistical problem for chapter 2 here.
• 14:24
  Oh, the. Visibility issue, yes.
• 14:27
  To patch a system in three days, or even to perform forensic
• 14:30
  triage on it, you have to actually know what exists in the
• 14:33
  first place. And from what Dan Ricci points
• 14:36
  out in the Clarity Nexus Insights we reviewed, most plant
• 14:39
  managers are operating off of glorified Excel spreadsheets IT.
• 14:42
  Is terrifyingly common. How do you?
• 14:44
  Patch a phantom asset you can't secure What you don't know you
• 14:47
  have you don't. And that is why visibility is
• 14:50
  the absolute foundation of any cybersecurity program.
• 14:53
  But in operational technology, visibility is an absolute
• 14:57
  minefield. Dan.
• 14:58
  Ricci, the ICS advisory founder. He makes a phenomenal point in
• 15:01
  that Cyber Scoop white paper. He argues that a simple list of
• 15:04
  assets, like that spreadsheet that has been updated since
• 15:07
  2024, is not a true asset inventory.
• 15:10
  Not even close. He calls it an illusion of
• 15:12
  control. And he's.
• 15:13
  Right. A static list is obsolete the
• 15:15
  moment you hit save. Think about a modern factory
• 15:18
  floor. Devices are constantly being
• 15:21
  swapped out for maintenance, IP addresses shift around,
• 15:24
  contractors walk in and plug temporary diagnostic laptops
• 15:28
  into the switches. It's a living.
• 15:29
  Environment it. Is a true inventory must be
• 15:32
  dynamic, organized, constantly updated and physically
• 15:36
  validated. It has to tell you not just what
• 15:38
  the asset is, say a Siemens PLC, but its current firmware
• 15:42
  version, its configuration state, its exact physical
• 15:45
  location in cabinet 4 on the plant floor, and its network
• 15:48
  dependencies like. Who is it talking to and who is
• 15:51
  allowed to talk to it exactly? So I was looking at traditional
• 15:54
  IT scanning tools, right? The stuff you'd use on a
• 15:56
  corporate network to build that dynamic inventory and think
• 15:59
  about applying them to an OT environment.
• 16:01
  And it feels like it feels like trying to secure a hospital by
• 16:05
  sending a heavily armed SWAT team to aggressively shake every
• 16:09
  single door, window and medical monitor to see if they're
• 16:12
  locked, that is. A great way to picture it, yes.
• 16:14
  You might find the unlocked doors, but you're going to
• 16:16
  accidentally unplug a bunch of life support machines in the
• 16:19
  process because you're being so aggressive.
• 16:21
  The. Mechanics behind that analogy
• 16:23
  are shockingly accurate, and the NIST OT Security Guide actually
• 16:26
  provides the historical receipts to validate it.
• 16:29
  Tell me about the historical stuff because this blew my mind
• 16:32
  so. Active scanning using a tool
• 16:34
  like an NAMP scan to map a network is standard practice in
• 16:39
  corporate ITIT systems use robust protocols like TCPIP, and
• 16:44
  they are designed to gracefully drop unexpected or malformed
• 16:48
  packets. If a server gets a weird ping,
• 16:50
  it just ignores it, right? It's built for the noise of the
• 16:53
  Internet, but. OT devices, especially legacy
• 16:55
  ones built 20 years ago, were never designed to handle
• 16:58
  unexpected network negotiation. They're.
• 17:00
  Incredibly fragile. Extremely.
• 17:02
  Fragile legacy device devices often run on primitive real time
• 17:05
  protocols like Modbus or DNP 3. They have tiny microcontrollers
• 17:10
  with very, very limited memory buffers.
• 17:12
  NIST documents this horrifying accidental event where a well
• 17:16
  meaning security team was performing a routine ping sweep
• 17:19
  on an active scatter network. And a ping sweep is basically
• 17:23
  just a network tool sending a message asking hello are you
• 17:26
  there to every IP address and arrange right?
• 17:29
  Yes. Just an ICMP echo request.
• 17:32
  Just a. Digital knock on the door.
• 17:34
  Nothing malicious at all, right? But this specific network they
• 17:37
  were scanning controlled A dormant 9 foot robotic
• 17:41
  manufacturing arm. Oh.
• 17:42
  Man the. Controller for the ARM received
• 17:45
  the ping sweep because it lacked a modern networking stack.
• 17:48
  It literally didn't know how to process the unexpected protocol
• 17:51
  negotiation. The tiny memory buffer on the
• 17:54
  controller overflowed, the CPU spiked to 100% and the system
• 17:58
  seized up just from a. Ping just from.
• 18:00
  A ping and when these industrial systems seize, they often
• 18:03
  default to a hard coded fail safe state.
• 18:05
  In this case, for whatever reason, the fail safe was an
• 18:07
  active state. Wait.
• 18:09
  The dormant 9 foot robotic arm suddenly swung 180° at full
• 18:14
  speed across the factory floor. Oh my.
• 18:16
  God, yeah. If anyone had been standing in
• 18:18
  the operational envelope of that arm, it wouldn't have been a
• 18:21
  cyber security incident. It would have been a needle
• 18:23
  industrial accident. Exactly the kinetic impact of a
• 18:26
  digital scan. And that wasn't an isolated
• 18:29
  incident. NIST details another event where
• 18:31
  a similar active scan was run on an ICS network in a
• 18:35
  semiconductor fabrication plant. Those.
• 18:37
  Clean rooms where they make microchips, yes.
• 18:40
  The system controlling the precise creation of the
• 18:42
  integrated circuits couldn't handle the malformed packets and
• 18:46
  it hung. It just locked up mid process
• 18:49
  because the environment in a fab plant must be perfectly
• 18:52
  controlled down to the chemical level.
• 18:54
  That single millisecond of digital hesitation resulted in
• 18:58
  the destruction of $50,000 worth of silicon wafers, instantly 50.
• 19:02
  Grand gone in a millisecond just because someone in IT wanted to
• 19:05
  know what IP addresses were active on the network IT.
• 19:08
  Highlights the massive cultural divide between IT and OT so.
• 19:11
  If you are the CI. So what does this all mean for
• 19:14
  visibility? If we can't actively scan the
• 19:16
  network without breaking the blast furnace or swinging A
• 19:18
  robotic arm around, how do we actually find the assets?
• 19:21
  We can patch them in three days in.
• 19:23
  Industrial scenarios passive monitoring is the required
• 19:26
  foundation. Instead of actively
• 19:27
  interrogating the devices, you use network tap PS or span ports
• 19:32
  to just listen to the ambient traffic I want.
• 19:34
  To focus on the mechanics of that for a second, just to be
• 19:36
  clear, if you're utilizing a span port, a switched port
• 19:41
  analyzer, you are configuring a network switch to basically take
• 19:45
  a photocopy of all the traffic flowing through it and send that
• 19:48
  copy to a monitoring tool, correct?
• 19:50
  And a. Tap or test Access point is a
• 19:52
  physical hardware device you literally insert in line into
• 19:55
  the network cable that physically copies the light or
• 19:58
  the electrical signal. Yes.
• 20:00
  In both cases, your monitoring tool is completely out of band.
• 20:03
  It is silently listening to the conversations already happening
• 20:06
  on the network between the Hmis and the PLCS.
• 20:09
  It's eavesdropping. Safe eavesdropping.
• 20:11
  It isn't generating a single packet of new traffic.
• 20:13
  It's completely Safe Life passive collection.
• 20:17
  The tool analyzes the packet headers and the payloads to
• 20:19
  establish baselines for asset behaviors.
• 20:22
  It identifies connected devices and it can even deduce firmware
• 20:25
  versions without ever directly touching the sensitive
• 20:27
  controllers. OK, so problem solved.
• 20:30
  We just sit back, listen to the span port, and build our dynamic
• 20:33
  inventory safely. If.
• 20:34
  Only it were that simple There's.
• 20:36
  Always a catch. Always.
• 20:37
  A catch. And this is where we have to
• 20:39
  look at the massive caveat introduced in the Enhancing OT
• 20:43
  Visibility White Paper. Passive monitoring is fantastic,
• 20:48
  but it relies on a fundamental law of physics.
• 20:51
  It only works if the assets are actually talking.
• 20:54
  If a device never speaks, a passive listener will never know
• 20:57
  it's there. So you.
• 20:58
  Have a massive blind spot for quiet devices.
• 21:01
  Significant blind spots. The white paper details a
• 21:04
  specific exception in the oil and gas industry that breaks the
• 21:07
  passive model entirely. Let's hear it.
• 21:09
  Imagine a critical remote programmable logic controller
• 21:13
  out on a distant unmanned oil pipeline in the middle of the
• 21:17
  desert somewhere. It communicates back to the
• 21:19
  central Scatta system over a high latency satellite link or a
• 21:23
  cellular modem, which is. Expensive and slow?
• 21:25
  Exactly. To save bandwidth and power, the
• 21:28
  engineers configured it to only communicate very very
• 21:30
  infrequently. Maybe it only sends a tiny
• 21:32
  telemetry ping once a week, or only when a pressure threshold
• 21:35
  is breached. So your.
• 21:36
  Passive monitor might sit there on the network for six days
• 21:40
  listening and hear absolutely nothing, leading you to assume
• 21:44
  that IP address. Exactly.
• 21:46
  And even when it does communicate, those sporadic tiny
• 21:49
  telemetry pings rarely contain the deep metadata you need.
• 21:53
  They provide incomplete firmware data which makes vulnerable in
• 21:56
  assessments and that CISA 3 day patch prioritization mandate
• 22:00
  completely impossible. You.
• 22:02
  Can't patch a vulnerability if you don't know what firmware the
• 22:05
  remote pipeline is running, right?
• 22:06
  In those highly specific cases, passive monitoring simply isn't
• 22:10
  enough. Organizations have to use
• 22:12
  throttled, highly controlled active.
• 22:14
  They have to poke the. Bear very.
• 22:16
  Gently, they have to carefully query those remote PLCS using
• 22:20
  native industrial protocols to validate their critical asset
• 22:24
  firmware. It's a terrifying, delicate
• 22:26
  balancing act between the risk of not knowing what's on your
• 22:29
  network and the risk of breaking the system by asking it a
• 22:32
  question and. When do you usually take that
• 22:34
  risk? You wait for the system to be
• 22:36
  turned off, Which brings us directly to Chapter 3, the
• 22:39
  concept of planned downtime. If you have to do that active
• 22:43
  collection, or if you have to apply that three day critical
• 22:46
  patch that requires a system reboot, you're usually waiting
• 22:49
  for a scheduled maintenance window to do it safely, Yes.
• 22:53
  The maintenance window, traditionally viewed by plant
• 22:56
  operators and engineers as the only safe harbor in the
• 22:59
  operational storm. The one.
• 23:00
  Time you can breathe the. One time you can actually touch
• 23:03
  the systems without risking A kinetic disaster, but the.
• 23:07
  Intelligence we reviewed from the Shieldworks report
• 23:09
  completely shatters that illusion.
• 23:11
  They reframe maintenance windows not a safe downtime, but is
• 23:15
  highly vulnerable, exceptionally high risk cybersecurity
• 23:18
  operations and if you. Look at the mechanics of how a
• 23:21
  plant actually operates during a turn around.
• 23:23
  They are absolutely right to reframe it.
• 23:25
  I mean I. Was laughing darkly reading this
• 23:27
  because it is the ultimate tragic irony.
• 23:30
  The exact time organizations bring their systems down to fix
• 23:34
  them, to patch them, to secure them against attackers, is the
• 23:36
  exact time they are most completely blind to attacks.
• 23:40
  It is. A massive operational blind spot
• 23:43
  think about the physical reality of what happens during a planned
• 23:47
  plant shutdown or a major turn around it's.
• 23:50
  Chaos it is. Organized chaos.
• 23:53
  Hundreds of pieces of equipment are taken offline, thousands of
• 23:56
  alarms are disabled so the sirens aren't blaring constantly
• 23:59
  in the control room. Systems are being rebooted left
• 24:02
  and right, and 3rd party vendors, often dozens of
• 24:05
  contractors who don't normally work there, are brought in with
• 24:08
  their own laptops to update proprietary machinery.
• 24:11
  The perimeter. Defenses are completely
• 24:12
  bypassed. It's essentially Swiss cheese.
• 24:14
  It's. Worse than just the perimeter
• 24:16
  being porous, the internal visibility is intentionally
• 24:19
  completely compromised. The Shield Works report details
• 24:23
  how OT monitoring tools, the very passive listening systems
• 24:26
  meant to detect anomalies we just talked about, are often
• 24:29
  literally powered down during these windows they turn.
• 24:32
  Off the security cameras, basically or.
• 24:34
  If they are left on, they are manually tuned into maintenance
• 24:37
  mode and the engineers do this because if they don't, the
• 24:40
  monitoring tool will trigger 10,000 critical alerts as
• 24:43
  contractors start disconnecting cables and flashing firmware
• 24:46
  and. Even if a security analyst is
• 24:47
  sitting there watching the screen, they are blinded by the
• 24:51
  sheer volume of legitimate, highly disruptive
• 24:54
  reconfiguration traffic. Precisely.
• 24:57
  It creates an impenetrable digital fog.
• 25:00
  Under the cover of that fog, malicious activities can be
• 25:03
  initiated without triggering a single alarm because anomalous
• 25:07
  behavior suddenly looks completely indistinguishable
• 25:10
  from normal maintenance behavior.
• 25:12
  The report. Gives some terrifying concrete
• 25:14
  examples of how adversaries exploit this exact dynamic.
• 25:19
  Let's look at the 2020 European manufacturing ransomware attack,
• 25:22
  that is. A classic example of exploiting
• 25:24
  the maintenance perimeter. In that incident, the
• 25:26
  manufacturing plant was undergoing a scheduled
• 25:28
  maintenance weekend. The local engineers needed an
• 25:31
  external vendor to dial in and patch some specialized systems
• 25:34
  remotely because the. Vendor is in another country,
• 25:36
  probably right? So they temporarily opened up
• 25:38
  Remote Desktop Protocol RDP sessions right through the
• 25:41
  firewall specifically for the vendor, but they.
• 25:43
  Left them open and unmonitored. Yes, adversaries constantly scan
• 25:48
  the Internet for open RDP ports. They found those temporary,
• 25:52
  unmonitored sessions, brute force the credentials, exploited
• 25:56
  the access, and deployed ransomware across the entire
• 25:59
  manufacturing environment. The plant was down for weeks.
• 26:02
  The mechanical lesson there is that temporary access, even for
• 26:06
  legitimate maintenance, must never, ever be left unmonitored.
• 26:10
  But it. Gets even stealthier than
• 26:12
  opportunistic ransomware, doesn't it?
• 26:14
  Look at the mechanics of the 2022 Asian refinery apartment
• 26:17
  attack that. 1 was an absolute master class in exploiting the
• 26:21
  maintenance window fog of war, Advanced persistent threat
• 26:24
  actors, an unattributed, highly sophisticated state sponsored
• 26:28
  group had likely been in the IT network for a while just.
• 26:31
  Watching. Just waiting patiently for a
• 26:33
  planned multi week shutdown at the refinery.
• 26:35
  They knew that during the shutdown, OT engineers would be
• 26:38
  rushing to update systems and flash new firmware onto the
• 26:41
  PLCS, so they. Hid their malicious update
• 26:43
  inside the legitimate update process.
• 26:45
  Exactly. While the engineers were pushing
• 26:48
  updates, the adversary slipped in and inserted persistence deep
• 26:52
  into the PLC firmware itself. They wrote to the EPROM of the
• 26:56
  controllers. Wait.
• 26:57
  Eprom deep in the actual physical hardware running the
• 27:00
  refining process. You can't just run an antivirus
• 27:03
  scan to find that, no. You can't.
• 27:04
  And because it happened during a maintenance window when firmware
• 27:08
  hashes were supposed to be changing, the passive monitoring
• 27:11
  tools didn't flag the new code as malicious.
• 27:14
  We just thought. It was part of the update, it
• 27:15
  just recorded it. Is another maintenance event.
• 27:18
  When the plant came back online and steady state operations
• 27:20
  resumed, the adversaries had a quiet, deeply embedded foothold
• 27:25
  that went completely undetected for months.
• 27:28
  They had total control over the refining process.
• 27:31
  OK, I have. To push back on behalf of the
• 27:32
  operators here, if you're a plant manager listening to this,
• 27:35
  what are you supposed to do? You've got compressed timelines.
• 27:38
  You have immense financial pressure from the board to
• 27:40
  minimize downtime. I mean in a large refinery every
• 27:44
  hour you are offline cost millions of dollars, dollars in
• 27:47
  laws production oh the. Pressure is unimaginable.
• 27:49
  You have. Dozens of vendors screaming to
• 27:52
  get their updates installed so they can move to the next job.
• 27:55
  The engineers are exhausted working double shifts.
• 27:57
  They're naturally going to cut corners just to get the physical
• 28:01
  plant running again. You can't expect them to act
• 28:04
  like forensic analysts it. Is an incredibly high pressure
• 28:08
  environment without question. The stress is immense, but the
• 28:11
  shield work solution is non negotiable.
• 28:14
  If you want the plant to remain secure in this threat landscape,
• 28:18
  you have to fundamentally reframe the culture around the
• 28:21
  maintenance window. It cannot be viewed as relaxing
• 28:23
  downtime. It's not a.
• 28:24
  Break. No, it must be treated as a high
• 28:27
  risk pre incident scenario with heightened situational
• 28:30
  awareness. So.
• 28:32
  You treat the maintenance window with the exact same security
• 28:34
  posture as you would an active cyber incident response.
• 28:38
  Exactly. Every single contractor's laptop
• 28:40
  needs to be scanned before it connects.
• 28:42
  Every temporary remote access port needs a strict time bound
• 28:46
  security baseline, but the real mechanical key to solving this
• 28:49
  is in the aftermath. Shieldworks outlines mandatory
• 28:52
  post maintenance threat hunting because.
• 28:55
  You can't just flip the switch back on you.
• 28:56
  Cannot just flip the switch, turn the plant back on and hope
• 29:00
  for the best. You have to.
• 29:01
  Clear the fog you have. To systematically validate every
• 29:04
  single configuration change against the planned work orders.
• 29:08
  If your monitoring tool says a PLC firmware changed on Tuesday
• 29:12
  at 3:00 AM, you cross reference that with the ticketing system
• 29:15
  you check. The receipts you.
• 29:16
  Verify who changed it, why they changed it, and you validate the
• 29:20
  cryptographic hash of the new firmware.
• 29:22
  If there is a change without a ticket, you treat it as a
• 29:25
  hostile breach immediately post maintenance threat hunting isn't
• 29:29
  optional anymore. It is the only way to ensure you
• 29:32
  haven't brought an attacker back online with you.
• 29:35
  And if that fog isn't cleared, if an attacker does slip through
• 29:39
  during maintenance, the consequences bleed rapidly out
• 29:42
  of the network cables and into the physical world.
• 29:45
  Which leads us to Chapter 4. Let's move to the structural
• 29:48
  defenses that are supposed to prevent that.
• 29:51
  We've talked about the vulnerabilities in time, the
• 29:53
  maintenance windows, but what about the structural
• 29:55
  vulnerabilities of the network architecture itself?
• 29:57
  This. Is where we look to the IEC
• 29:59
  624443 standard, which is globally recognized as the
• 30:04
  foundational framework for industrial cybersecurity
• 30:07
  architecture. The core philosophy of the
• 30:10
  standard relies heavily on 2 specific concepts, zones and
• 30:14
  conduits, right? Let's define the mechanics of
• 30:16
  these clearly for anyone who might be designing a network.
• 30:20
  A zone is a logical or physical grouping of assets that share
• 30:24
  common security requirements. Correct.
• 30:27
  For example, you wouldn't put a general purpose engineering
• 30:29
  workstation in the same zone as a critical safety controller.
• 30:33
  You might group all the human machine interfaces, the Hmis
• 30:36
  which are the touch screens the operators use, into one zone.
• 30:40
  Then you group the critical safety programmable logic
• 30:42
  controllers into to an entirely different zone because the
• 30:45
  safety controllers require a massively higher level of
• 30:47
  protection. The stakes.
• 30:48
  Are different exactly. If the HMI goes down, you lose
• 30:51
  visibility. If the safety controller goes
• 30:53
  down, the plant explodes. They cannot share the same
• 30:56
  security posture and the. Conduit is the logical grouping
• 30:59
  of communication channels that protects the data flowing
• 31:02
  between those zones. It's the pipe connecting the
• 31:05
  rooms. But how does this actually work
• 31:07
  in practice? Like if I'm a network architect
• 31:09
  building this out, what is the physical or logical reality of a
• 31:13
  conduit? It's not just a cable, right?
• 31:16
  No, it depends entirely on the security levels or SLS of the
• 31:21
  zones being connected. Let's contrast two different
• 31:24
  implementations to make it concrete.
• 31:26
  If you have a conduit connecting 2 zones that share the exact
• 31:29
  same security level, say connecting one group of Hmis to
• 31:33
  another group of Hmis, you might implement A transparent logical
• 31:36
  gateway. OK, what?
• 31:37
  Does that do this? Acts like a protected pipe.
• 31:40
  Data flows relatively freely, but the conduit ensures the
• 31:43
  integrity and confidentiality of that data in transit.
• 31:46
  Perhaps using IP SEC encryption or host identity protocols to
• 31:50
  ensure no one is spoofing the packets in the middle OK.
• 31:52
  That makes sense for peer-to-peer lateral traffic,
• 31:55
  but what if I need to connect the corporate IT network, which
• 31:57
  is a fundamentally low security level full of phishing emails
• 32:00
  and web browsers, to the planned floor controllers, which are a
• 32:03
  high security level in? That scenario, a transparent
• 32:06
  gateway, is completely unacceptable for a transition
• 32:09
  between a low security zone and a high security zone.
• 32:13
  Your conduit must implement an application Layer gateway or
• 32:17
  ALG. How's.
• 32:18
  That different from a normal firewall this.
• 32:20
  Isn't just a firewall looking at IP addresses and port numbers.
• 32:24
  It acts as a highly restrictive deep packet inspection filter.
• 32:28
  It cracks open the industrial protocol packets.
• 32:31
  It inspects the traffic at the application layer ensuring only
• 32:35
  specific pre approved commands are allowed to cross the
• 32:38
  boundary. So it actually understands the
• 32:40
  industrial language. Exactly.
• 32:41
  If the protocol is Modbus, the ALG knows the difference between
• 32:45
  a read coil command and a write coil command.
• 32:47
  Oh, that's clever. It can be configured to allow
• 32:50
  the corporate IT network to send a read command to check the
• 32:52
  temperature of a sensor for a dashboard, but it will
• 32:55
  absolutely drop and block any write command attempting to
• 32:58
  change the temperature setpoint. The conduit strictly enforces
• 33:01
  the segmentation at the command level.
• 33:03
  And when that strict segmentation fails, or when a
• 33:06
  network is built without it entirely, it truly is a matter
• 33:11
  of kinetic destruction. The historical postmortems from
• 33:14
  NIST and Anissa that we reviewed are chilling because they show
• 33:19
  the physics of what happens when zones bleed together.
• 33:22
  Take the 2014 German steel mill attack that.
• 33:25
  Incident remains A watershed moment in OT security because of
• 33:28
  the sheer scale of the physical damage walk.
• 33:31
  Us through it. Hackers breached the corporate
• 33:33
  Business Network first, most likely through a standal social
• 33:37
  engineering tactic or a spear phishing e-mail targeting an
• 33:40
  office worker. But because the network wasn't
• 33:43
  properly segmented with strict conduits and AL GS, the
• 33:46
  attackers didn't hit a wall. The.
• 33:47
  Zones bled together. They were.
• 33:49
  Able to pivot directly from the IT network down into the
• 33:52
  production networks running the steel mill and.
• 33:54
  Once they were in the production network, they didn't just steal
• 33:56
  data, right? They manipulated the control
• 33:58
  system. They.
• 33:59
  Systematically disrupted the control components of a massive
• 34:02
  blast furnace. And we have to understand the
• 34:05
  physics of a blast furnace to really understand the
• 34:07
  consequence here. Yeah, paint.
• 34:08
  A picture for us A. Blast furnace operates
• 34:10
  continuously at temperatures exceeding 2000°C, smelting iron
• 34:15
  ore. If you stop the cooling systems
• 34:18
  or if you halt the extraction process abruptly, the molten
• 34:21
  iron inside will begin to cool and solidify.
• 34:24
  It turns. Into a 500 ton solid block of
• 34:27
  iron inside the furnace, yes. What the industry calls a
• 34:30
  salamander, it essentially destroys the furnace from the
• 34:33
  inside out, requiring months of explosives and jackhammers to
• 34:37
  clear. Wow.
• 34:38
  The attackers disrupted the control systems to such a severe
• 34:41
  degree agree that when the plant operators realized something was
• 34:44
  wrong and tried to perform a standard controlled shutdown of
• 34:48
  the blast furnace to save it, the system failed to respond the
• 34:51
  brakes. Were out the.
• 34:52
  Automated safety sequences were corrupted.
• 34:54
  The furnace could not be shut down properly, resulting in
• 34:57
  massive catastrophic physical destruction to the facility.
• 35:00
  The. Malicious code physically broke
• 35:02
  the mill. And then there's the 2017 Triton
• 35:05
  attack on the Saudi petrochemical facility.
• 35:08
  This one is particularly insidious, arguably more evil,
• 35:11
  because it didn't just target the main process controllers, it
• 35:14
  specifically targeted the safety systems.
• 35:16
  Triton is the absolute nightmare scenario for any industrial
• 35:20
  engineer in any hazardous industrial environment.
• 35:23
  You have the distributed control system, the DCS, which runs the
• 35:27
  day-to-day process, but completely separate from that
• 35:30
  you have a safety instrumented system, or SIS.
• 35:32
  Like a physical watchdog, the SIS operates independently of
• 35:36
  the main process controllers. It monitors the physical state
• 35:39
  of the plant if the main process goes out of control.
• 35:42
  If pressure builds up too high in a pipe or temperatures spike
• 35:45
  towards an ignition point, the SIS acts as the autonomous last
• 35:49
  line of defense to trigger AN automatic safe shutdown before
• 35:53
  an explosion occurs, and the. Hackers bypassed the main
• 35:56
  process and specifically wrote custom malware to compromise
• 36:00
  that safety system. They wanted to blind the
• 36:02
  watchdog, yes. The attackers compromised an
• 36:05
  engineering workstation and used a proprietary protocol to deploy
• 36:08
  malicious software that reprogrammed the Triconex safety
• 36:12
  controllers. These controllers use a
• 36:13
  sophisticated triple redundant processor architecture.
• 36:16
  How does? That work 3.
• 36:17
  Independent microprocessors all process the same data and vote
• 36:21
  on the outcome. If two vote that the plant is
• 36:24
  safe and one votes it isn't, the two outvote the one.
• 36:27
  It's designed for extreme physical reliability.
• 36:30
  But the malware tried to rewrite the logic on all three processes
• 36:35
  at the same time exactly. Fortunately for the facility,
• 36:38
  the attackers made a slight coding error.
• 36:40
  This error caused a timing mismatch in the execution of the
• 36:43
  malicious code across the redundant processors.
• 36:46
  So they. Tripped over their own
• 36:47
  shoelaces, they. Did the processors realized they
• 36:50
  were out of sync, triggering a hard coded fail safe state?
• 36:53
  This caused the plant to autonomously shut down safely
• 36:56
  before the attackers could execute their ultimate payload.
• 36:59
  Thank God, but the intent of Triton was undeniable.
• 37:02
  Quietly disable the safety system, wait, and then cause a
• 37:05
  catastrophic physical explosion on the main process.
• 37:08
  Without thesis there to stop it, it's just.
• 37:11
  It's pure malice and the vulnerability isn't isolated to
• 37:15
  manufacturing or oil and gas. ENISA, the European Union agency
• 37:20
  for cybersecurity, outlines similar highly destructive
• 37:23
  threat scenarios for National Railway networks.
• 37:26
  We aren't just talking about a single plan anymore, we are
• 37:29
  talking about sprawling infrastructure.
• 37:31
  The. Anisa report details how a lack
• 37:33
  of strict segmentation could allow malware introduced via a
• 37:37
  simple phishing e-mail in a rail operator's corporate office to
• 37:41
  propagate into the operational technology systems of the rail
• 37:44
  network itself. The signalling.
• 37:45
  Systems. They describe scenarios where
• 37:48
  adversaries could manipulate the traffic supervising systems,
• 37:51
  injecting malicious shell code that interrupts the automatic
• 37:53
  train protection or ATP signals. The ATP is what physically
• 37:58
  prevents trains from colliding if you disrupt it.
• 38:01
  The fail safe is an urgent emergency halt of all train
• 38:04
  traffic across the network. You.
• 38:05
  Completely paralyze the logistics of a nation instantly.
• 38:08
  And. Even if they don't hit the
• 38:09
  trains directly with malware. And Lisa points out that
• 38:12
  unsegmented Internet exposed systems are incredibly
• 38:16
  vulnerable to brute force disruption, they highlight
• 38:19
  scenarios where bot Nets launch massive distributed denial of
• 38:23
  service, or D das attacks against Internet facing railway
• 38:27
  ticketing, scheduling and routing services.
• 38:29
  You don't actually have to derail a train to shut down a
• 38:32
  transit system. You just have to take down the
• 38:34
  digital systems that allow people to buy a ticket or know
• 38:36
  what track the train is on. Exactly the kinetic impact is
• 38:40
  achieved through digital congestion, but.
• 38:42
  There's an entirely different layer to this structural
• 38:45
  vulnerability, one that is brilliantly highlighted by the
• 38:47
  engineering experts at E Tech Group.
• 38:50
  In cybersecurity, we focus so heavily on digital malware, the
• 38:53
  Russian hackers, the RAN somewhere gangs, that we
• 38:56
  completely forget the physical network infrastructure itself is
• 38:59
  often failing invisibly. Yes.
• 39:01
  Invisible physical network failures E tech group points out
• 39:05
  that unexpected network slowdowns, packet loss, and
• 39:09
  these ghost communication errors on the plant floor often have
• 39:12
  absolutely nothing to do with the cyber attack and everything
• 39:15
  to do with the hostility of the physical environment because.
• 39:17
  It's not a pristine server room. Industrial environments are
• 39:21
  fundamentally hostile to delicate electronics.
• 39:24
  A corporate server room is climate controlled and pristine,
• 39:28
  a factory floor is not. E Tech Group notes how ambient
• 39:32
  heat from blast furnaces and microscopic factory dust slowly
• 39:35
  degrade network switch components and choke cooling
• 39:38
  fans over years and the vibration.
• 39:40
  Furthermore, the constant low frequency vibration of heavy
• 39:43
  machinery like massive stamping presses or turbines literally
• 39:47
  loosens the physical Ethernet and serial cable terminations in
• 39:50
  the network racks and. Electrical noise, right?
• 39:52
  The electromagnetic interference or EMI.
• 39:54
  EMI is a massive factor. When you have massive industrial
• 39:58
  motors running nearby, they generate huge electromagnetic
• 40:01
  fields. This interference can actually
• 40:03
  corrupt the digital signals, the ones and reads traveling inside
• 40:07
  the unshielded network cables. It causes cyclic redundancy
• 40:11
  check or CRC errors. Packets get dropped or malformed
• 40:14
  it. Creates this vicious feedback
• 40:16
  loop of failure it. Does because the.
• 40:18
  Physical degradation of the cable causes digital packet
• 40:21
  loss. The packet loss decreed to the
• 40:23
  reliability of the control loos between the HMI and the LC.
• 40:26
  The degraded control LOO leads to unstable physical processes,
• 40:30
  which causes more vibration, which loosens the cable further.
• 40:34
  It proves that in OT, digital and physical reliability are
• 40:37
  completely, deeply intertwined. A loose vibrating wire can mimic
• 40:41
  the exact symptoms of a cyberattack, and a sophisticated
• 40:44
  cyberattack can exploit the instability caused by a loose
• 40:46
  wire, which. Makes governing and defending
• 40:49
  these systems incredibly complex the defender is fighting both
• 40:52
  nation States and the laws of physics simultaneously and.
• 40:55
  It's only going to get harder as we add more autonomy.
• 40:59
  Let's move into the final chapter of our discussion
• 41:01
  exploring the future of OT defense.
• 41:04
  We've talked about AI as an offensive weapon, you know the
• 41:07
  negative time to exploit. But AI is also being rapidly
• 41:10
  integrated into OT defense and daily operations.
• 41:13
  The. SIASA OT Zero Trust Guidance
• 41:15
  tackles this emerging reality head on.
• 41:18
  Utilities, manufacturers and logistics companies are heavily
• 41:21
  integrating artificial intelligence into their
• 41:23
  environments to manage processes that are simply too complex for
• 41:27
  human operators to optimize. Like what?
• 41:30
  They are using machine learning models for predictive
• 41:32
  maintenance, analyzing massive data sets of acoustic vibrations
• 41:35
  to predict exactly when a pump will fail days before it
• 41:38
  happens. They're also using AI for
• 41:41
  autonomous load balancing on power grids, dynamically routing
• 41:44
  electricity based on real time demand and.
• 41:47
  Crucially, they're using AI based security platforms to
• 41:50
  monitor the network traffic we talked about earlier.
• 41:52
  The AI listens to the span ports and learns what is normal so it
• 41:55
  can detect the anomalies. But there is a terrifying
• 41:59
  concept buried in the CISA guidance regarding these
• 42:02
  defensive AIS and it's called Model Drift model.
• 42:05
  Drift is arguably one of the most insidious invisible threats
• 42:09
  in modern OT environments. To understand it, you have to
• 42:13
  understand how machine learning actually works.
• 42:16
  Break it down. For us.
• 42:17
  When you train an AI model to detect anomalies on a factory
• 42:20
  floor, you train it on a baseline of normal data.
• 42:24
  The AI builds a complex mathematical model of what the
• 42:27
  network and the physical processes look like when
• 42:29
  everything is fine. But in a physical industrial
• 42:32
  environment, normal is always changing, right?
• 42:35
  Because of the physical degradation we just talked
• 42:36
  about. Exactly.
• 42:37
  Equipment ages. Bearings slowly wear down,
• 42:41
  causing their acoustic frequency to shift slightly.
• 42:43
  Seasonal temperature changes alter the fluid dynamics of a
• 42:46
  pipeline, changing the pressure signatures.
• 42:49
  A new contractor installs a slightly different brand of
• 42:52
  sensor. It's never.
• 42:53
  Static day. By day, month by month, the real
• 42:56
  world operating conditions slowly diverge from the original
• 42:59
  static distribution of the AI's training data.
• 43:01
  So the. AI's accuracy degrades.
• 43:04
  It's looking for a baseline that no longer physically exists.
• 43:08
  Think of model drift like the magnetic deviation on a ship's
• 43:11
  compass. That's a.
• 43:12
  Great way to think about it the. Compass isn't broken.
• 43:15
  The needle is still pointing to what it mathematically
• 43:17
  calculates as N, but because the magnetic field around the ship
• 43:21
  has slowly changed over the years, the compass is north is
• 43:24
  now 3° off from true north. If the navigator doesn't
• 43:28
  recalibrate it, you smoothly sail your ship straight into a
• 43:31
  coral reef, all while the sophisticated instrument panel
• 43:35
  tells you your heading is perfectly fine that.
• 43:37
  Is a phenomenal analogy because it highlights the core danger.
• 43:41
  It degrades silently when an AI model drifts, it doesn't throw a
• 43:45
  blue screen of death. It doesn't sound an alarm.
• 43:48
  The degradation manifests as incrementally less reliable
• 43:51
  predictions. So you.
• 43:52
  Don't even know it's failing an. OT security platform whose
• 43:55
  underlying ML model has drifted might continue to report 99%
• 43:59
  confidence scores to the dashboard, assuring the CI so
• 44:02
  that the network is secure. Meanwhile, it's actual
• 44:05
  mathematical ability to detect a novel subtle threat has
• 44:08
  completely eroded. It's.
• 44:10
  A total false sense of security. You are completely defenseless
• 44:14
  against the sophisticated attacker, and absolutely no one
• 44:17
  in the control room knows it. So how can an organization
• 44:20
  possibly govern this kind of invisible mathematical threat?
• 44:24
  You can't just install an antivirus program for model
• 44:27
  drift CISA. Provides a very specific
• 44:29
  structural recommendation here. True cross functional fluency.
• 44:33
  Historically the IT security team, the OT engineering team on
• 44:37
  the plant floor, and the data science team building the AI
• 44:41
  operated in completely separate silos.
• 44:43
  They didn't even. Talk to each other.
• 44:44
  They didn't. Speak each other's languages,
• 44:46
  and frankly, they often didn't like each other because.
• 44:48
  IT cares about data confidentiality and patching, OT
• 44:51
  cares about physical safety and keeping the line running at all
• 44:55
  costs, and data science cares about algorithms, rhythmic
• 44:57
  accuracy, and data purity. They have fundamentally
• 45:00
  conflicting priorities. Exactly.
• 45:03
  But CESA states that to govern AI and OT, these groups must
• 45:06
  collaborate at the policy level continuously, not just when a
• 45:09
  crisis hits. During an incident response, the
• 45:12
  communication loop must be tight.
• 45:14
  Give me an. Example of what that looks like
• 45:16
  for. Example, if the OT engineer
• 45:18
  replaces a physical pump with a slightly different model, they
• 45:21
  have to proactively tell the data scientist.
• 45:24
  The data scientist then knows they must retrain the AI model
• 45:27
  on the new acoustic baseline of the new pump and if.
• 45:30
  They don't if they. Don't communicate this.
• 45:33
  The security team's dashboard is going to suddenly get flooded
• 45:36
  with false positives because the AI thinks the new pump sound is
• 45:40
  an anomaly. Or much worse, the AI model
• 45:43
  adjusts itself poorly and creates false negatives,
• 45:46
  accepting a Stuxnet style physical manipulation as normal
• 45:50
  wear and tear. Which is.
• 45:51
  Terrifying. Genuine.
• 45:52
  Continuous cross functional fluency is the only way
• 45:56
  governance frameworks actually shape the physical risk posture
• 45:59
  of the plant. It's about.
• 46:00
  Breaking down the cultural silos to build actual operational
• 46:04
  resilience, and that concept of resilience brings us to the
• 46:07
  Nordal philosophy outlined in our final source text.
• 46:11
  Nordal emphasizes a very hard, uncomfortable truth for the
• 46:14
  cybersecurity industry. Perfection is impossible.
• 46:16
  It is. The hardest truth to accept, but
• 46:18
  the most necessary. If the exploit window is
• 46:21
  negative because of AI weaponization, if active
• 46:24
  scanning breaks the factory, if maintenance windows are a
• 46:27
  vulnerable fog of war, and if your AI defensive model is
• 46:31
  silently drifting out of accuracy, you mathematically
• 46:34
  cannot guarantee prevention. The.
• 46:36
  Fortress will be breached the. Fortress will absolutely be
• 46:39
  breached. So what?
• 46:40
  Does resilience actually mean then if not keeping the bad guys
• 46:44
  out? Resilience means accepting the
• 46:46
  inevitability of the threat. It means assuming the adversary
• 46:49
  will bypass the perimeter and designing your architecture to
• 46:52
  minimize the physical impact of that breach and ensuring you can
• 46:55
  recover operations quickly and safely.
• 46:57
  Which? Ties back to the logs it ties.
• 46:59
  Directly back to the Clarity Insights and the IEC 62443
• 47:04
  standards specifically requirement 6.1 audit log
• 47:07
  accessibility because. If you don't have robust read
• 47:10
  only tamper proof audit logs aggregated in the central
• 47:13
  location, you cannot perform the forensic triage that sees some
• 47:17
  mandates in that three day window.
• 47:18
  Precisely. You can't clear the fog of the
• 47:21
  maintenance window without logs. If APLC goes down and you don't
• 47:25
  have the logs to tell you whether it was a network packet
• 47:27
  overflow, a loose cable, or a Russian APT injecting firmware,
• 47:32
  you don't know what happened, You're blind, And if you don't
• 47:34
  know what happened, you don't know how to recover safely.
• 47:36
  You might reboot a compromise system and trigger the
• 47:39
  destructive payload. Resilience is built on the
• 47:42
  assumption of failure. It is continuous AI assisted
• 47:46
  threat hunting. It is rigorous physical and
• 47:49
  logical segmentation with application layer gateways.
• 47:52
  It is cross functional communication between the
• 47:55
  engineer and the data scientist, and above all is an unwavering
• 47:59
  obsessive commitment to visibility.
• 48:00
  It is. A completely different paradigm.
• 48:02
  We aren't building an impenetrable static fortress
• 48:05
  anymore. We are building a ship with
• 48:07
  enough intelligent, watertight compartments to stay afloat,
• 48:10
  navigate the damage, and keep moving forward even after it
• 48:13
  takes a direct torpedo hit. That's.
• 48:15
  All for this episode of Cybersecurity Under Pressure
• 48:18
  attacks. Real essence.
• 48:19
  Check your systems, check your assumptions, and maybe, just
• 48:22
  maybe question whether the AI model or the air gap you trusted
• 48:25
  yesterday is still reliable today.
• 48:27
  See you in the next chapter.