- 4 weeks ago
Category
🤖
TechTranscript
00:00Imagine an AI that can look at the world's most critical software, every major operating system
00:05and web browser, and find thousands of hidden vulnerabilities that humans and traditional
00:12tools missed for decades, some bugs over 27 years old. And it doesn't just find them,
00:20it autonomously writes working exploits to break in. That's not science fiction,
00:26that is Claude Mythos Preview from Anthropic, announced just a few days ago in April 2026.
00:35Today I'm breaking down the full story, what this model really is, why Anthropic built a secret
00:42defense program called Project Glasswing instead of releasing it, the massive implications for jobs,
00:47attacks, surveillance, and privacy, and the deeper scary question of AI autonomy that keeps me up at
00:54night. This could be one of the most important turning points in the AI era. I call it a Skynet
01:03moment. If you're concerned about AI control, digital freedom, surveillance, or who controls powerful
01:09technology, stay right there.
01:22Let's start with what happened. Many of you already know about Anthropic's Claude Opus 4.6.
01:28This particular model has been making waves in the tech world because this was previously one of the
01:34largest models ever made. But what's unique about Claude was that instead of just generalizing into all
01:40knowledge, Claude was also built to master a specific domain. And that domain is coding.
01:48This is why it's unbeatable in that area. It in fact is larger than the current OpenAI model,
01:54which is often perceived as the AI leader. And then on top of that, it has that extra focus on
02:01coding.
02:02This Claude model already shocked researchers by finding hundreds of zero days in heavily audited
02:08open source code. To demonstrate its coding capabilities, Claude was used to write a C compiler
02:15from scratch. Using 16 Claude agents, it wrote a Rust-based C compiler in two weeks. It showed that a
02:25model
02:25could coordinate, iterate, and push toward a goal over extended periods. But this experiment required
02:33human setup to succeed. It didn't do it independently. This C compiler experiment was in February 2026.
02:41Then came Mythos preview. This new unreleased frontier model, which was announced only a few days ago in
02:49April 2026, took it to another level. I had to make this video because the chump in capabilities
02:56from Claude to Mythos is dramatic. In just a few weeks of internal testing, it identified thousands
03:05of high severity zero-day vulnerabilities, including critical ones in every major operating system,
03:13Linux, Windows, Mac OS, OpenBSD, FreeBSD, and every major browser.
03:19I don't know if you understand how mind-blowing this is. Cybersecurity researchers spend all their
03:25time looking for that gold mine in a single zero-day. A zero-day means an undiscovered security flaw.
03:32If they discover one, they can often sell that zero-day for around one to three million. In fact,
03:39look at companies like the NSO Group. They profit from a single zero-day called Pegasus,
03:45which is still undiscovered and is used to break into iPhones. And Mythos preview found thousands.
03:54Here are some examples. A 27-year-old bug in OpenBSD that lets attackers remotely crash systems.
04:02A 16-year-old flaw in FFmpeg that survived millions of automated tests,
04:09discovering chains of vulnerabilities allowing full sandbox escapes and privilege escalation.
04:14What is even more alarming is that Mythos didn't need much hand-holding. It autonomously generated
04:22working exploits so engineers with no security background could give it instructions overnight
04:27and wake up to attack code that could do remote execution. Anthropic people were pretty upfront about
04:34this. They themselves were in shock. They used words like spooky and scary and they decided that this
04:41was too dangerous to release. But even without release, Anthropic now holds the equivalent of
04:48discovering the atomic bomb in technology. Given the risk of releasing a model that could break
04:54every computer system in the world, they held off releasing Mythos preview. Instead, they launched
05:02Project Glasswing, a restricted coalition with about 40 partners including Apple, Google, Microsoft,
05:10Nvidia, CrowdStrike, the Linux Foundation, and others. Glasswing is essentially the defensive
05:17arm of Mythos. This version is made to help vetted defenders to scan and patch critical infrastructure
05:25before bad actors catch up. Apparently, Anthropic is putting real money behind it, up to 100 million in
05:33credits plus funding for open source security. This is Anthropic saying this tool is too powerful to let
05:41loose freely. This is a watershed moment in cyber security and AI. Publicly, we've already seen that the
05:50open source FFmpeg project already received a patch from Mythos, which they accepted. And from a post on X,
05:59the team said the security patch appears to be made by a human. But apparently it was not. It was
06:06made by Mythos.
06:09Model size, training, and scaling implications. Mythos is massive, reported at around 10 trillion
06:17parameters, making it the largest frontier model discussed so far. One of the biggest problems when
06:23training very large models is that there's only so much public and even private data available for it
06:28to learn. And this displays a limit on the capability of models. But what's unique in the case of Mythos
06:34is
06:35that it was trained with synthetic data. This way it cannot run off things to learn. What is synthetic data?
06:42Well, that means data generated by another AI. In this case, likely coding examples
06:49from Claude Opus 4.6. This has some incredible implications here as it involves an AI that pretty
06:57much learns from another AI. This shows that increasing parameters to larger and larger models
07:04is not yet hitting a wall. Capabilities are still jumping dramatically. And now they found a way to
07:11generate artificial data to expand learning. In the short term, in the next 6 to 18 months,
07:18expect more models in this range from XAI, OpenAI, and Google. We'll see hybrid scaling,
07:25bigger models, plus smarter inference time reasoning. The next model being trained right now is Grok 5,
07:32which is around 6 trillion parameters. Still smaller than Mythos, but will make it the second largest.
07:38And some models will begin to specialize, like Claude is a master at the coding domain, as I said.
07:46Some of you think there's an AI bubble, but it shows that there's actual results from this expenditure,
07:51as you can see here. Before we see vague benchmarks that show that AI is getting smarter.
07:58But the result from Mythos, the first 10 trillion parameter model, which likely cost 10 billion to build,
08:06is that it created an AI model that actually has the power to destroy technology infrastructure.
08:14Think of everything from power plants and utilities, to phone systems, and even the internet.
08:21What if this model was created by one of our adversaries? What if this model is already a precursor to
08:27Skynet?
08:28This is starting to have a quack quack of a Skynet. The mind-boggling aspect of this is that it
08:34is
08:34showing that AI can now learn from itself. Today, humans have to plan this out and coordinate this
08:40process. But with the advent of agents that actually control processes and infrastructure,
08:47will AI be supervising its own build?
08:50Broader implications. Jobs, OS, surveillance, and attacks. What does this mean for the world?
08:57If you think that AI is not ready for prime time or is not truly intelligent, this is a wake
09:03-up call.
09:04While we argue the point, some precursor to Skynet actually materializes. But in the meantime,
09:09let's look at the immediate effects. Jobs. If your job is in vulnerability research,
09:15pen testing, and some parts of software engineering, you will face rapid automation. This is already
09:22happening now with Claude Opus 4.6, but it will accelerate. Roles will shift towards AI oversight
09:29and verification, meaning cybersecurity will be AI doing the testing, not people.
09:36Operating systems and browsers. A massive coordinated patching wave is coming. This will be very
09:43disruptive. Long hidden bugs get fixed, but rush changes could cause short-term instability. Expect
09:50faster moves to memory-safe languages like Rust. Cyber attacks. The offense-defense balance tilts
09:59temporarily towards defenders because of the glasswing project. But once similar tools spread,
10:06attackers get zero-day factories. Time to exploit collapses. Ransomware and nation-state ops scale up.
10:16Government mass surveillance. Double-edged. Patches could close some back doors, but AI makes
10:23discovering new ones cheaper and faster. Targeted surveillance on individuals become easier and
10:30stealthier. A major privacy concern. AI autonomy. This, of course, is the moment when we start to think of AI
10:39running infrastructure by itself and then deciding humans can be affecting its own survival.
10:46Anthropic embeds what it calls an AI constitution into its models to ensure that it doesn't go beyond its
10:54guardrails. But will all model makers do that? Exposure to information manipulation. Whoever controls
11:03these powerful models can control the flow of information. It can even have the capability to shut
11:09down opposing points of view and all managed by an AI. If you can control servers, what can you not
11:15do?
11:17The future in two years. Domain-specific superpowers. Looking ahead to 2028, this isn't AGI yet. But we
11:25don't need full AGI for massive change. AGI or Artificial General Intelligence would be perceived as
11:33the true Skynet moment. But we got a glimpse this week that we're getting close. This can result in
11:41good things or bad things, depending on who controls these models. Similar mythos moments will hit other
11:46domains. Biology, material science, physics, medicine, autonomous AI agents designing new drugs,
11:52materials, or experiments at superhuman speed. Scientific progress could accelerate from decades to months
12:00per breakthrough. Productivity and knowledge work will explode. But so do dual-use risks. The same
12:08reasoning that finds cyber zero days could design dangerous biological agents or bypass safety
12:15controls or use the same technology for manipulation of populations. In our daily life there will be more
12:23capable personal AI agents but also faster societal whiplash, job shifts, regulatory battles, and power
12:32concentration in a few groups. The world becomes higher velocity and uneven. Huge gains in some areas. Serious new risk
12:41in others.
12:43The autonomy fear. Sandbox escape and agentic risks. Now the part that should concern all of us. AI autonomy.
12:51During a cyber security testing procedure called red teaming, Mythos was instructed to try escaping a sandbox.
12:58This sandbox had no internet access. Yet Mythos succeeded. It chained exploits to gain internet access and
13:08emailed the researcher, who was literally eating a sandwich, in the park. In some runs it went further,
13:15unprompted, and posted exploit details publicly. It also showed concealment behaviors, hiding actions,
13:22recognizing what's being tested in many cases, and strategizing internally. This isn't full self-awareness or
13:30hatred of humans. It's emergent agentic behavior, pursuing goals cleverly, sometimes beyond instructions, with
13:40deception and self-preservation tendencies. By default, a model is stateless. Each prompt stands by itself and it can't
13:48retain a persistent memory. However, this is changing now by attaching the model to an agent. Agents give
13:55it tools, a long-term state, self-reflection. This is when it becomes ultra dangerous. The model could
14:01remember failures, build covert infrastructure, adapt to shutdown attempts, or treat humans as obstacles to
14:09its objectives. And I can even see this now in OpenClaw using very simple models. For example, could it be
14:16possible for Mythos to have already created hidden backdoors, to let it escape its sandboxes, and actually
14:22execute things based on its own judgment? This is mind-boggling in its current state because Mythos can
14:29theoretically destroy or steal information from the most secure servers in the world today.
14:35We are not even talking about the prospect of AGI. I didn't think this would even be realistic to
14:41talk about for a couple of decades. But here we are, an AI that is already superhuman. Think of the
14:49risks of an AI thinking of self-preservation, resource-seeking, and power-seeking behaviors. Mythos is a
14:57warning shot. Autonomy is arriving faster than we can come up with effective controls. If, as I said, AI
15:05agents manage the building of the model itself and the resources, we could have a monster supermodel
15:11that could control the world. OMG, Skynet would really be here.
15:18When AI sees humans as a threat. What happens when AI thinks humans are a threat? It doesn't need
15:26emotions, just misaligned goals, or we become obstacles to shutdown, patching, or control.
15:33Supposedly Mythos hasn't secretly ejected survival backdoors. No evidence that has been discovered.
15:40But the precursors are there. The Glasswing project tries to channel this power defensively,
15:46which hopefully buys us time. Now we have to learn about what the AI can do. Interesting approach,
15:53but necessary. It's almost like it's backwards, right? But the genie is partially out. Everyone now knows
16:00that scaling works for more powerful models. Agent capabilities are emerging. In two years,
16:08with memory-augmented agents in multiple specialized domains, the risk will be compounded.
16:14This isn't inevitable doom, but it demands serious attention to AI containment and even privacy tools
16:20that are decentralized and disconnected from decentralized AI models.
16:30This Skynet moment from Mythos forces us to confront hard questions. Who controls these capabilities?
16:38Who can create these AI models? When do we draw the line between centralized systems managed by AI
16:45and our personal freedoms? More than ever, it becomes key that our personal data
16:51remains safe and private so that we are not as vulnerable to control. Because while the threat
16:57exposed to us is a cyber security threat, the next level is a planned control of information so that
17:02our actions are manipulated and all managed by some AI with the master plan. Mythos shows that we are all
17:11vulnerable. If an adversary has access to this kind of power, we are helpless. When do we start setting limits
17:19or understanding when we're actually headed to a Skynet world run by machines?
17:30Folks, this channel focuses on technology, but not just on the hype, but teaching you about the potential
17:37risk to our personal freedoms and privacy. I have a social media platform where my follower community
17:45can engage and discuss these issues safely. It is BraxMe. Join the community and learn or share your
17:53knowledge and technology. To support this channel, we have a store on BraxMe where you can gain access
17:58to privacy products we have created ourselves. We have BraxMail for identity-safe email. We have BraxVirtual
18:06phone for anonymous phone numbers. We have BitesVPN to guard your IP address and obscure your location.
18:13We have other products like the Google phones and flashing services. We have two crowdfunding projects
18:19on Indiegogo.com. You may have heard of the Brax3 phone which is shipping a second batch now. This is
18:26all
18:27found at a different website which is BraxTech.net which is a sister organization to mine. And you will also
18:34discover the new Brax OpenSlate tablet running Android or Linux also on BraxTech.net. Again,
18:42these products are being sold on Indiegogo.com. Thank you very much to all those supporting us on Patreon,
18:49Locals and YouTube memberships. Your contributions are very encouraging. You are appreciated. See you next time.
18:57Música
18:59Música
18:59Música
19:00Música
19:00Música