Anthropic's Most Dangerous Model Reveal Raises Concerns Nobody's Talking About | Ep 3

00;00;00;09 - 00;00;22;27

Imagine you build the most powerful AI model your company has ever made. It's so good at one thing in particular that you decide you're just not going to sell it. Not on the API, not in the app, not to anyone. You lock it in a vault and hand the keys to a group of governments and fortune 500 companies.

00;00;23;00 - 00;00;31;21

This is not hypothetical. This is a real model, and it's called Claude Mythos. It exists right now and you can't use it.

00;00;31;21 - 00;00;52;16

So today we're answering three questions. One, what is mythos, actually? Two why would a company that lives and dies by shipping models refuse to ship its best one yet? And three, the question I haven't seen anyone say out loud is the timing of all of this a little too convenient?

00;00;52;20 - 00;00;55;15

Because anthropic, after all, is about to go public.

00;00;55;18 - 00;00;56;28

Let's get into it.

00;00;56;28 - 00;01;16;17

Okay. Facts first. No hype. For years, anthropic line up followed a simple pattern. Haiku for speed. Sonnet for balance. Opus for raw power. Mythos broke that pattern. It's not just a bigger opus, it's positioned as a whole new tier sitting above opus.

00;01;16;20 - 00;01;25;08

The existence of the model leaked first back on March 26th, reportedly through draft blog posts that got out before they were supposed to.

00;01;25;12 - 00;01;34;06

Anthropic confirmed it to fortune, and on April 7th, they made it official with a model called Claude. Mythos preview.

00;01;34;06 - 00;01;35;26

Here's where it gets wild.

00;01;35;28 - 00;01;44;15

On paper, mythos is just a general purpose model, strong across math, long context reasoning, and software engineering.

00;01;44;15 - 00;02;01;27

But the headline capability, the one that made everyone stand up,

00;01;48;22 - 00;02;02;00

is cybersecurity and anthropic own framing is that this was almost an accident. They pushed coding and reasoning, and out the other end came a model that's frighteningly good at finding holes in software.

00;02;02;00 - 00;02;06;08

How good? Let me give you the numbers, because they're the whole story.

00;02;06;08 - 00;02;07;16

In its first month.

00;02;07;17 - 00;02;13;03

Mythos reportedly uncovered over 10,000 high or critical severity vulnerabilities.

00;02;13;03 - 00;02;20;10

Anthropic Red team says it found flaws in every major operating system and every major web browser.

00;02;20;16 - 00;02;31;05

here's the part that should make your stomach drop. Over 99% of the vulnerabilities it found were not yet patched, meaning they were live, real, and exploitable.

00;02;31;10 - 00;02;32;18

the prompt to do it

00;02;32;22 - 00;02;37;24

essentially, please find me a security vulnerability in this program.

00;02;36;24 - 00;02;37;20

That's it.

00;02;37;27 - 00;02;44;09

They also reported that engineers with no security training could use it to produce complete working exploits.

00;02;44;13 - 00;02;52;09

Now on the spec sheet it's just a 1 million token context window. 128 K output adaptive reasoning model.

00;02;52;12 - 00;02;53;28

specs aren't the point.

00;02;53;28 - 00;03;02;21

The point is, this is the first model where the gap between find the bug and weaponize the bug basically collapsed to a single sentence.

00;03;02;26 - 00;03;05;09

if it's that powerful, why not sell it?

00;03;05;09 - 00;03;07;24

Every other lab races to ship.

00;03;07;24 - 00;03;11;27

Anthropic answer is something called Project Glass Wing.

00;03;11;27 - 00;03;23;25

Instead of a public release, they spun up a consortium, a closed club that gets to use mythos to find and fix vulnerabilities in critical software before the attackers can.

00;03;23;25 - 00;03;33;05

The reported members are a who's who. Identity security firm Okta, Korean giants like Samsung, SK Hynix and SK Telecom,

00;03;33;05 - 00;03;35;18

the military alliance NATO

00;03;35;22 - 00;03;39;08

and the EU cybersecurity agency. And.

00;03;39;11 - 00;03;45;07

It's reportedly already scaled to critical infrastructure across 15 plus countries.

00;03;45;07 - 00;03;52;07

The logic is what they call defensive acceleration. If a model this good at offense is going to exist

00;03;52;07 - 00;04;07;19

in anthropic openly says rivals will have one soon, such as OpenAI, which already has shown a cybersecurity focused model in response. Then you better get it into the hands of defenders first, so they can patch up the world before the bad guys catch up.

00;04;07;19 - 00;04;17;25

It's a genuinely interesting safety argument. Don't democratize the most dangerous thing you've ever built. Give it only to the people fixing the locks, not picking them.

00;04;17;27 - 00;04;30;01

Oh, and the one detail I love. The model briefly showed up in the public version of Cloud Code, listed as Claude Mythos. One preview with a toggle to turn it on. Then it vanished.

00;04;30;07 - 00;04;38;27

So somewhere someone is clearly preparing it for a wider rollout. The vault door isn't welded shut, it's just closed for now.

00;04;38;29 - 00;04;49;03

Now here's where I want to take the headphones off the press release and actually have you think. Watch this timeline. March 26th. Mythos leaks

00;04;49;03 - 00;04;50;07

April 7th.

00;04;50;07 - 00;04;56;24

Anthropic discloses the too dangerous to release superweapon and launches Glass Wing with NATO and the EU.

00;04;56;29 - 00;05;10;03

Late May, anthropic closes a whopping $65 billion funding round at an insane $965 billion valuation, passing OpenAI for the very first time.

00;05;10;06 - 00;05;14;04

Early June, anthropic confidentially files for its IPO.

00;05;14;10 - 00;05;24;18

So let me ask the uncomfortable question is Claude Mythos partly a narrative engineered to prop up the valuation right before anthropic goes public?

00;05;24;18 - 00;05;38;15

Let me Steelman the skeptic, because the case isn't crazy.

00;05;28;00 - 00;05;38;15

One the IPO story. Anthropic is selling. Wall Street is coding in cybersecurity dominance. That's literally what the analysts say. Lifted them above OpenAI.

00;05;38;17 - 00;05;47;08

what's the most dramatic possible proof of cybersecurity dominance? Well, a model so powerful you claim you can't even sell it.

00;05;47;08 - 00;05;50;04

That's not a product. That's the making of a legend.

00;05;50;04 - 00;05;56;07

It's unfalsifiable marketing. You can't test the claims because you can't access the model

00;05;56;09 - 00;05;58;24

to look at who validates it.

00;05;58;24 - 00;06;05;27

NATO, the EU, Samsung you couldn't buy more credible IPO roadshow logos if you tried.

00;06;05;27 - 00;06;13;01

Governments trust us with the dangerous one is the single best line a frontier lab could walk into a public offering.

00;06;13;05 - 00;06;21;19

Three the convenient leak A leak that forces you to disclose your scariest, most impressive capability ever.

00;06;21;21 - 00;06;24;10

About ten weeks before you filed a go, public

00;06;24;10 - 00;06;29;27

skeptics will say the most flattering possible accident is rarely actually an accident.

00;06;29;27 - 00;06;36;07

Okay, now let me argue the other side, because I think the cynical take is a little too easy.

00;06;36;07 - 00;06;37;18

Counterpoint one

00;06;37;18 - 00;06;41;25

the cybersecurity establishment is terrible at playing along with fake hype.

00;06;41;25 - 00;06;45;16

The Alan Turing's Institute Security Center analyzed this.

00;06;45;17 - 00;07;13;27

If the vulnerability claims were vapor, the security research community would have absolutely shredded them publicly within a week. They didn't. Counterpoint two there's real downside here. Telling the world you built a model that finds unpatched zero days in every browser is not a clean flex. It's an admission that the same capability is coming for everyone, and it invites regulators, lawsuits, and national security scrutiny.

00;07;13;28 - 00;07;16;29

Right. As you're trying to look like a stable public company.

00;07;17;02 - 00;07;19;24

That's a strange thing to fabricate on purpose.

00;07;19;24 - 00;07;23;13

Counterpoint three the leak cuts against the stunt theory.

00;07;23;13 - 00;07;27;10

You don't usually plan to lose control of the rollout of your own narrative.

00;07;27;10 - 00;07;34;10

A genuine leak that forced an early and messy disclosure looks less like choreography and more like damage control.

00;07;34;10 - 00;07;35;28

So where does that leave us?

00;07;35;28 - 00;08;00;18

Probably in the middle. Mythos is very likely a real model with real, very scary capabilities, and anthropic is absolutely sophisticated enough to understand that a two dangerous to release model is the best IPO story money can't buy, and then lean into it.

00;07;54;10 - 00;08;00;21

Both things can be true. The capability can be genuine, and the framing can be strategic.

00;08;00;21 - 00;08;06;08

The thing to watch here, and this is the tell, is what happens after the IPO prices.

00;08;06;12 - 00;08;18;17

If mythos quietly becomes a paid tier in clawed code six months from now, the two dangerous to sell framing was always temporary. Watch the vault door. It tells you more than the press release.

00;08;18;17 - 00;08;20;16

So zoom out.

00;08;20;17 - 00;08;22;16

What's the actual lesson here?

00;08;22;20 - 00;08;24;09

We've crossed a line.

00;08;24;12 - 00;08;34;07

We now live in a world where the most capable AI models aren't measured by how well they write your emails. They're measured by whether they're too dangerous to be released at all.

00;08;34;10 - 00;08;41;15

That's a different era. The frontier isn't just a chatbot anymore. It's a security weapon with a safety lock on it.

00;08;41;19 - 00;08;44;11

here's the part that matters for regular people.

00;08;44;14 - 00;08;51;22

the best AI gets locked behind consortiums of governments and trillion dollar companies, access becomes the whole game.

00;08;51;24 - 00;09;00;21

The gap isn't people who use AI and people who don't. It's between people who get the powerful stuff and the people who get the leftovers.

00;09;00;26 - 00;09;06;15

That's exactly why model access shouldn't be a velvet rope. It should be a vending machine.

00;09;06;15 - 00;09;12;21

If this episode made you think and made you a little suspicious in a healthy way. Hit subscribe.

00;09;12;21 - 00;09;19;07

We pay you in tokens to learn, and we're just getting started. I'm Alex Smith and this is super confident AI