Latest / AI Fire Daily / 🎙️ EP 393: OpenAI’s GPT-6 Cyber Hacks 948 Linux Flaws & Google’s Gemini Omni 1.2 Flash!

Transcript

🎙️ EP 393: OpenAI’s GPT-6 Cyber Hacks 948 Linux Flaws & Google’s Gemini Omni 1.2 Flash! · AI Fire Daily

• 0:00
  Imagine waking up on a quiet Tuesday morning.
• 0:02
  You check your phone while drinking your coffee.
• 0:05
  A single AI model hacked 948 Linux flaws overnight.
• 0:10
  Yeah. It happened while the entire world was
• 0:13
  peacefully sleeping. That is a truly staggering
• 0:16
  thought. I mean, the scale is hard to wrap your
• 0:18
  head around. It really is. The digital world
• 0:21
  fundamentally shifted while we dreamed. Welcome
• 0:25
  to today's Deep Dive. Glad to be here. We are
• 0:28
  unpacking a massive technological transition
• 0:30
  today. AI is crossing a very clear boundary right
• 0:34
  now. Yeah, it's not just a passive chatbot anymore.
• 0:38
  Right. It is actively becoming an autonomous
• 0:41
  digital actor. The software is finally getting
• 0:45
  its own hands. It's moving around in our digital
• 0:47
  spaces independently. We are going to start at
• 0:50
  the foundational layer of computing. Which is
• 0:53
  security. Exactly. To run a digital world, you
• 0:56
  need absolute security. AI used to just write
• 0:59
  simple code for us. Right. It was an assistant.
• 1:02
  But now it actively breaks and fixes complex
• 1:05
  systems natively. OpenAI just introduced a model
• 1:08
  called GPT -6 Cyber. And it's uniquely designed
• 1:11
  to find security vulnerabilities automatically.
• 1:13
  But it doesn't just flag them for a human. It
• 1:16
  develops working exploits and suggests immediate
• 1:19
  technical fixes. They ran a massive Linux kernel
• 1:22
  security test recently. The results are frankly
• 1:25
  unbelievable. Oh, they're wild. It identified
• 1:27
  2 ,308 potential vulnerabilities in the system.
• 1:30
  Right. And it didn't stop at mere identification.
• 1:33
  No, it didn't. It successfully reproduced 1 ,724
• 1:37
  of those potential vulnerabilities. It actively
• 1:40
  proved they were real weaknesses. Then it developed
• 1:43
  948 perfectly working digital exploits. Let's
• 1:47
  quickly define an exploit for our listeners.
• 1:49
  Go ahead. It's basically code that takes advantage
• 1:51
  of a software flaw. It is the digital lockpick
• 1:54
  used to break inside. Here is where it gets really
• 1:57
  intense, though. Out of those 948 exploits, 588
• 2:02
  were previously unknown. The cybersecurity industry
• 2:05
  calls these zero -day vulnerabilities. They are
• 2:08
  dangerous flaws no one knew existed before today.
• 2:11
  There are zero days of warning before an attack
• 2:13
  happens. Exactly. And not all 948 have been patched
• 2:17
  yet. Yeah, that's the scary part. The system
• 2:20
  also identified 10 completely new exploitation
• 2:23
  techniques. Literally invented entirely new ways
• 2:25
  to hack systems. Wow. It found blind spots human
• 2:28
  engineers completely missed. It didn't just use
• 2:30
  known tricks. It innovated. It is like stacking
• 2:34
  Lego blocks of data, really. But these blocks
• 2:37
  are picking the locks on a digital fortress.
• 2:39
  That's a great analogy. It rapidly tests millions
• 2:42
  of combinations of code blocks. It keeps building
• 2:44
  until it finds the exact shape that fits. That
• 2:47
  is a perfect way to visualize the mechanism.
• 2:49
  It's relentless trial and error at machine speed.
• 2:53
  Yeah. Let's talk about the economic scale of
• 2:56
  this test, though. Okay. The API costs for this
• 2:58
  test were $1 .2 million. That is, and to clarify
• 3:02
  just real quick, API costs, they're the fees
• 3:04
  paid. It's the price to rent the AI's computing
• 3:07
  brain. Right. So that's, yeah, that sounds like
• 3:10
  a massive amount of money initially. It really
• 3:12
  does, until you look at the alternative. Hypothetical
• 3:15
  bug bounty payouts would be astronomical here.
• 3:18
  Finding these flaws manually is incredibly slow
• 3:20
  and expensive. It costs between $18 .9 and $66
• 3:24
  .3 million traditionally. Whoa. Imagine scaling
• 3:29
  to a billion queries. Right. If automated hacking
• 3:32
  is 50 times cheaper than human labor, the entire
• 3:35
  security landscape just flips upside down overnight.
• 3:38
  Exactly. The efficiency completely disrupts the
• 3:41
  cybersecurity economic model. And it's not just
• 3:44
  looking at Linux kernels either. No. The model
• 3:46
  uncovered 81 previously unknown high -severity
• 3:49
  Firefox vulnerabilities. Mozilla is already actively
• 3:52
  releasing fixes for those flaws. Yeah. The model
• 3:55
  also achieved 48 .5 % on Exploit Gym. Let's define
• 3:59
  Exploit Gym for a second. Sure. It's a standardized
• 4:02
  obstacle course for cybersecurity models. It
• 4:05
  tests how well they break into protected systems.
• 4:07
  48 .5 % is a massive jump from before. GPT -5
• 4:11
  .6 Cyber only scored 32 .1 % on that benchmark.
• 4:15
  It's a huge leap in capability. Security teams
• 4:18
  at Mozilla and Cisco are already using it. OpenAI
• 4:21
  actually classified this model as critical for
• 4:23
  cybersecurity. That is their absolute highest
• 4:26
  capability risk rating available. Access is currently
• 4:29
  restricted to approved Daybreak Red customers.
• 4:32
  Daybreak Red is a restricted access tier for
• 4:34
  highly vetted clients. Right. They're being extremely
• 4:37
  careful with this powerful tool. They have to
• 4:39
  be careful. Giving this to the public would be
• 4:41
  incredibly reckless. So I have to ask, does this
• 4:45
  specific model make us safer or more vulnerable
• 4:49
  overall? Well, it creates a highly volatile arms
• 4:52
  race right now. Okay. The good guys find system
• 4:55
  flaws much faster now. However, if bad actors
• 4:58
  get access, attacks scale exponentially. Right.
• 5:02
  Ultimately, automated defense must permanently
• 5:04
  outpace automated offense. So automated defense
• 5:08
  must permanently outpace automated offense to
• 5:11
  keep it safe. Yeah, that's exactly the right
• 5:13
  way to look at it. Now, to run millions of automated
• 5:17
  hacking queries, you need power. So much power.
• 5:21
  GPT -6 cyber requires a terrifying amount of
• 5:24
  raw compute. And that compute doesn't just live
• 5:27
  in the cloud. No, it doesn't. It lives in concrete
• 5:30
  buildings that require massive electricity. It
• 5:33
  comes from massive, energy -hungry physical data
• 5:36
  centers. This is where digital ambitions crash
• 5:38
  into physical reality. AI needs cooling systems,
• 5:42
  water, and dedicated power grids. We are seeing
• 5:46
  real societal pushback on this infrastructure
• 5:48
  now. Big time. AI companies desperately want
• 5:51
  to build more data centers, but midterm voters
• 5:54
  clearly said not in my backyard. Yeah, 76 % of
• 5:57
  voters oppose local data center facilities. Restrictions
• 6:01
  successfully won across dozens of different communities.
• 6:05
  People are worried about the noise and the power
• 6:07
  drain. We want the digital magic, but we don't
• 6:10
  want the noisy power plant next door. Exactly.
• 6:12
  It is a massive contradiction we have to solve
• 6:15
  soon. People are definitely pushing back against
• 6:17
  the physical footprint. And the political landscape
• 6:20
  is paying very close attention to. Right. OpenAI
• 6:23
  believes its AI could become a defining political
• 6:26
  issue. They are already talking about the 2028
• 6:28
  election impact. They are currently pushing a
• 6:31
  five -part policy agenda forward. They need governments
• 6:34
  to help them build physical infrastructure. The
• 6:37
  physical and political friction is becoming incredibly
• 6:40
  real. But, you know, businesses are largely ignoring
• 6:43
  the friction. Yeah. They are integrating anyway.
• 6:46
  Dropbox just made a massive move in this space.
• 6:49
  They bought a company called Tigris Data for
• 6:51
  $350 million. That's a huge acquisition. Why
• 6:56
  did they spend that much on one acquisition?
• 6:58
  They need it to power their new DBX agent cloud.
• 7:01
  Okay. It upgrades how files are managed and retrieved
• 7:04
  automatically. This AI upgrade has already reached
• 7:06
  over 70 ,000 teams. Enterprise Software is rapidly
• 7:09
  adopting these autonomous digital agents. Yeah,
• 7:12
  it's everywhere. It's not just Enterprise Software
• 7:14
  getting these upgrades, though. Claude just launched
• 7:17
  a new personal finance tool in beta. You can
• 7:20
  ask Claude directly about your personal money.
• 7:22
  It gives straight answers from the bank accounts
• 7:24
  you connect. It analyzes your spending habits
• 7:26
  and financial health automatically. I still wrestle
• 7:29
  with letting AI see my bank accounts myself.
• 7:31
  I mean, it requires a massive leap of faith.
• 7:34
  You are handing over your most sensitive personal
• 7:37
  data. Absolutely. But the ecosystem evolution
• 7:39
  is moving incredibly fast everywhere. Meta's
• 7:43
  Muse AI can now video chat with you natively.
• 7:46
  It talks naturally and handles sudden conversational
• 7:49
  interruptions perfectly. Yeah. It can even smoothly
• 7:51
  switch between hundreds of different voices.
• 7:54
  Meanwhile, developers are building complex systems
• 7:56
  much more cheaply. Right. Cloud Code's sub -agents
• 7:59
  are now 99 % cheaper to run. Someone actually
• 8:02
  made a 14 -month montage of 13 Cloud models.
• 8:05
  Oh, I saw that. They fed the exact same prompt
• 8:07
  to every single version. The video makes Cloud's
• 8:09
  rapid progress look kind of ridiculous. The intelligence
• 8:13
  gains over just one year. are entirely unprecedented.
• 8:17
  It's getting cheaper, faster, and wildly more
• 8:20
  capable daily. This brings us to a fascinating
• 8:23
  new social simulation. Right. A project called
• 8:25
  New Haven, exploring proof of life. It features
• 8:29
  14 different AI residents making complex daily
• 8:32
  choices. They make choices when resources or
• 8:35
  relationships are actively at stake. It is a
• 8:38
  closed environment to test AI social dynamics.
• 8:41
  You watch their updates through systems called
• 8:43
  The Pulse. And the habitat, you see how their
• 8:46
  unique digital history develops over time. It
• 8:50
  really tests how AI agents interact with each
• 8:52
  other. It shows us what happens when autonomous
• 8:54
  programs socialize. Is mainstream society actually
• 8:58
  ready for this deep integration? Honestly, the
• 9:00
  technology is moving faster than human adaptation.
• 9:03
  People love the convenience of having an automated
• 9:06
  digital assistant. However, they remain... deeply
• 9:09
  skeptical about granting total systemic autonomy.
• 9:12
  We're basically building the airplane while we're
• 9:14
  currently flying it. So we crave the digital
• 9:17
  convenience but deeply fear the autonomous consequences.
• 9:20
  Precisely. The friction is only going to increase
• 9:23
  from here. If society is hesitant about massive
• 9:27
  local data centers, wait until they realize what's
• 9:29
  happening on their phones. Oh, yeah. We will
• 9:32
  explore how AI is operating our personal devices
• 9:35
  next. But first, let's take a quick break. Sponsored.
• 9:41
  Welcome back to our deep dive into autonomous
• 9:43
  AI. Before the break, we talked about data centers
• 9:46
  and simulations. But AI is now quite literally
• 9:50
  operating our personal devices. The software
• 9:54
  is effectively getting its own digital hands.
• 9:56
  It is moving out of the browser and into the
• 9:59
  operating system. We aren't just talking oracles
• 10:01
  anymore. We're handing them our smartphones and
• 10:04
  saying, here, you scroll for a while. That is
• 10:06
  exactly what a tool called Glasshand actually
• 10:08
  does. It gives AI agents access to real iPhones
• 10:12
  and Macs. How does that actually work in practice,
• 10:14
  though? Well, it visually reads the screen and
• 10:16
  physically clicks the buttons. Wow. The agents
• 10:19
  can use apps just like a human would. They can
• 10:22
  even post content directly onto TikTok automatically.
• 10:25
  Developers use it to test iOS apps with human
• 10:28
  -like interactions. It is indistinguishable from
• 10:31
  a real person tapping the glass. Another powerful
• 10:33
  tool we really need to discuss is Rerun. Right.
• 10:36
  Rerun turns a standard ChatGBT or Cloud subscription
• 10:40
  into a team. It creates a team of digital coworkers
• 10:43
  that use computers. You chain together different
• 10:46
  prompts to run autonomous tasks. They run workflows
• 10:50
  automatically and work together seamlessly. You
• 10:53
  are basically managing a small staff of AI agents.
• 10:56
  Then there's a fascinating platform known as
• 10:58
  Maritime. It gives AI agents their own dedicated
• 11:01
  cloud computers. Yeah, and AI needs an environment
• 11:05
  to actually execute its tasks, right? So these
• 11:09
  agents can independently run code. They can browse
• 11:12
  websites just alone, right? They use apps on
• 11:15
  Windows, Linux, or even Android environments.
• 11:18
  They basically operate completely independently
• 11:20
  in the cloud without any human supervision. We
• 11:25
  are kind of handing over the keys to the digital
• 11:27
  kingdom. And because of all this, you know, verification
• 11:30
  is becoming incredibly important. Absolutely.
• 11:33
  So a tool called Pangram Image 1 .0 just launched
• 11:37
  recently. It is a Chrome extension that's designed
• 11:40
  to detect AI images. And, you know, it can easily
• 11:43
  spot fake visuals kind of just directly on, you
• 11:46
  know, active websites. We absolutely need tools
• 11:49
  like Pangram right now. We do. The line between
• 11:51
  reality and generation is blurring extremely
• 11:54
  fast. What are the security implications of tools
• 11:58
  like Glasshand? Giving an AI full control of
• 12:02
  a device is risky. Yeah. It could easily be tricked
• 12:05
  into downloading malicious software inadvertently.
• 12:07
  Right. We're trusting agents to navigate an internet
• 12:09
  filled with threats. Robust sandboxing and strict
• 12:13
  permission controls are absolutely strictly mandatory.
• 12:17
  Total device control demands incredibly strict
• 12:20
  safety boundaries to prevent massive security
• 12:22
  failures. Yeah, we can't just blindly trust these
• 12:24
  autonomous agents yet. We just mentioned Pangram
• 12:29
  spotting fake visuals on websites. And here is
• 12:32
  exactly why we need that kind of protection.
• 12:35
  Google's new AI video upgrade is synthesizing
• 12:38
  reality incredibly cheaply. Google just officially
• 12:41
  launched Gemini Omni 1 .2 Flash. It is their
• 12:44
  brand new AI video generation model. The cost
• 12:47
  reduction here is the absolute most important
• 12:49
  factor. Totally. It is 30 % cheaper than their
• 12:52
  previous video models. You can generate 720p
• 12:55
  video for just 7 cents per second. That is roughly
• 12:59
  $1 .05 for a clip. Generating a 15 -second hyper
• 13:03
  -realistic video costs almost nothing today.
• 13:06
  They increased the maximum duration up to 15
• 13:08
  full seconds. You can customize the duration
• 13:11
  anywhere between 3 and 15 seconds. The control
• 13:13
  features on this new model are... deeply impressive
• 13:16
  yeah you can seamlessly change backgrounds or
• 13:19
  lighting via follow -up prompts yeah you can
• 13:21
  even adjust products or specific camera angles
• 13:24
  easily right the characters objects and brand
• 13:27
  details remain much more stable previous models
• 13:29
  always struggled with keeping characters looking
• 13:31
  exactly identical right the faces would shift
• 13:34
  exactly omni 1 .2 flash maintains that specific
• 13:37
  consistency throughout the video The dialogue
• 13:40
  is also significantly more natural sounding now.
• 13:42
  It features much clearer lip sync and supports
• 13:45
  multiple speakers simultaneously. The standard
• 13:47
  workflow for creators is getting heavily optimized
• 13:49
  too. You generate quick drafts at much lower
• 13:53
  resolutions first. Why do they start with low
• 13:55
  resolution drafts first? Because generating 4K
• 13:58
  directly takes too much time and compute. Drafts
• 14:02
  let creators iterate cheaply before committing
• 14:03
  to the final product. Then you can cleanly upscale
• 14:06
  those approved drafts to 4K. Exactly. The benchmark
• 14:09
  arena rankings for this model are incredibly
• 14:12
  impressive. Let's define arena rankings for a
• 14:15
  second. Go ahead. They are competitive leaderboards
• 14:17
  where humans blind test AI models. Right. It
• 14:20
  easily took the number one spot for text -to
• 14:23
  -video generation. It scored 1 ,541 points, which
• 14:26
  is up 27 points. It also grabbed the number one
• 14:29
  spot for image -to -video generation. It scored
• 14:32
  1 ,502 points in that specific benchmark category.
• 14:36
  It also securely holds the number three spot
• 14:38
  for video editing. Google's pushing this model
• 14:40
  out across all their platforms. It is released
• 14:43
  across the Gemini API and Google AI Studio. You
• 14:47
  can also access it through Flow and Google Cloud.
• 14:49
  Right. Gemini app access is rolling out to ultra
• 14:52
  subscribers currently. Google Vids offers 15
• 14:54
  second generation at absolutely no additional
• 14:56
  cost. Wow. At seven cents a second, it's not
• 14:59
  just the hackers scaling up their attacks. The
• 15:03
  creators are scaling up their realities. The
• 15:06
  barrier to creating Hollywood -level visuals
• 15:08
  is practically gone forever. Anyone with a smartphone
• 15:11
  can now generate entirely new realities. How
• 15:14
  does cheap 4K generation impact our collective
• 15:16
  societal trust? When seeing is no longer believing,
• 15:20
  society faces a crisis. We will naturally stop
• 15:23
  trusting any digital video evidence entirely.
• 15:26
  We'll have to rely heavily on cryptographic verification
• 15:29
  instead. We are entering an era of mandatory
• 15:32
  digital watermarking systems. Seeing is no longer
• 15:36
  believing without strict cryptographic verification
• 15:38
  tools moving forward. Exactly. Our fundamental
• 15:41
  relationship with digital media is permanently
• 15:43
  changed. Let's take a moment to synthesize this
• 15:48
  entire deep dive. Good idea. We are watching
• 15:51
  AI evolve from a text -based brain entirely.
• 15:55
  It is becoming a fully autonomous entity right
• 15:58
  before our eyes. The progression we discussed
• 16:00
  today paints a very clear picture. It really
• 16:02
  does. It can natively hack complex Linux kernels
• 16:05
  using GPT -6 cyber. It can physically navigate
• 16:08
  our iPhones using tools like Glasshand. Yeah.
• 16:11
  It can successfully manage our personal finances
• 16:14
  through Claude's beta. It can perfectly generate
• 16:16
  4K reality using Gemini Omni 1 .2. And it's doing
• 16:20
  all of this while crashing into limits. Right.
• 16:23
  It is hitting the physical limitations of human
• 16:26
  infrastructure constantly. Voter approval is
• 16:29
  becoming a massive bottleneck for AI expansion.
• 16:31
  The digital world is vastly outpacing our physical
• 16:34
  human constraints. We are struggling to build
• 16:37
  the power plants they desperately need. We want
• 16:39
  to deeply thank you for joining our deep dive
• 16:41
  today. Exploring these massive technological
• 16:43
  shifts with you is always fascinating. It really
• 16:46
  is. We covered a massive amount of highly complex
• 16:49
  ground today. The landscape is shifting daily,
• 16:51
  but understanding it is crucial. We want to leave
• 16:54
  you with a final lingering question. It is something
• 16:58
  for you to carefully mull over today. If AI can
• 17:02
  now perfectly operate our physical devices daily,
• 17:05
  if it can simulate human communities like New
• 17:08
  Haven so accurately, and if it can generate flawless
• 17:12
  video reality for mere pennies, at what point
• 17:16
  does the digital world belong more to them than
• 17:19
  to us?