All right, we're going to do this again.

I hope I hope you're ready for this,

okay? Because since June 12th, we've

been working closely with the US

government to restore access to Claude

Mythos 5 and Fable 5. Gee, I can't

believe Anthropic lost it. I wonder I

wonder what caused that. Today, the

government has notified us that mythos

5, our strongest cyber security model,

can be redeployed to a set of US

organizations that operate and defend

critical infrastructure. Now, you're

probably thinking, well, wait a sec.

organizations that do critical

infrastructure. Haven't we seen this

before? Wasn't there something about

mythos and only going out to select

organizations? Yes, there was. It was

called Project Glass Wing. This was just

a couple months ago when Anthropic said

Mythos was in fact too dangerous and

then they released it to everybody and

then they pulled it back because of the

US government was told how dangerous it

was. Well, we're just replaying this

entire record again. We're back to

Project Glass Wing. There is the set of

the good guys, which apparently is Jeff

Bezos and Sundar and Mark Zuckerberg.

They're the good guys. They get access

again and the rest of us, we don't get

access. Now, I'm going to talk a lot

about safety, okay? Cuz here's the deal.

I think this is all fake. I think not

only is it fake, I think it's actually

hurting people. I actually have the

opposite opinion of Daario. Surprise,

surprise. Me with the opposite opinion.

Me with the opposite opinion of this guy

right here. Okay. We don't actually see

eye to eye. By the way, this is Daario

in 2023. Oh, we're going to play this

clip of him talking to the old

government. just really scaring them up.

Of course, before we begin, I'd like to

say thank you to the sponsors. One of

the more terrifying things you can do as

a company is to contract out software.

And I know a lot of companies are

contracting out their mobile

applications. The thing that makes this

so dangerous is that it's a blackbox

experience. You hope you get back

software that is complete, that is going

to be well tested, that ultimately will

be extensible as future needs require.

The last thing I want someone else to

say to Claude, "Hey, mobile app, no

mistakes." It does make mistakes. This

is why I trust the guys over at Infinite

Red. I have had many along talks with

the CEO about software quality and the

way to really build robust and good

systems and I love everything he has to

say and everything that I have seen with

my own eyes. The quality of software

that they build is very very good. All

right, so the safety argument is

actually pretty dang ridiculous that

we're having right now. This whole kind

of gated access to claw mythos. But I

just wanted to kind of add one more

quick point. If you don't know, right

now Chad GPT 5.6 is also in a limited

preview and the reason being for now at

the request of the US government, we're

starting with a limited preview among a

small group of trusted partners in

Codeex and the API. Sam finally has a

model that's also too dangerous. He's

not allowed to release it either. We got

to make sure we slowly roll the thing

out. And we're entering into an era in

which the government is now going to

have a larger and larger impact into

what is exactly happening. And this was

always Daario's goal. His goal was

always to have the government step in

and make regulation. He's been just

saying this non-stop. If you read his

writings, if you listen to his talks,

and his urging that we all need to slow

down. Well, here we are. It's finally

happening. His dream has been realized.

The frontier studios are all being

slowed down. But that's not really

actually making anything safer because

here's the deal. Have you heard of the

old model GLM? Well, GLM 5.2 apparently

beats Claude in some cyber security

benchmarks. At least in this article,

the security that's being ran is

insecure direct object reference. If you

don't know what that is, that's simply,

let's just pretend you have an endpoint

on your website that goes and grabs user

information. If I were to pass in any

user ID, you should probably go, hey,

are you that user? Oh, you're not? Hey,

you're not allowed to look at I'm sorry.

Hey, brother. Sorry. This is not you.

404 don't know who this is. Well, it

turns out that the fourth most common

bug on hacker one top vulnerability list

is in fact these right here. These oopsy

poopsies insecure direct object

references where people can just go onto

websites and get data if they just

simply have an ID. So Sam Gre went off

and developed a benchmark against these

type of vulnerabilities. And it turns

out things like Claude Code, Opus 46 and

48 did about a 28% to 37%. Jippidity

Codeex did about a 20% and GLM with no

harness, just the model itself did 39%

did better than both Claude Code and

their latest models and Jeypity 55. Now,

if you don't know, a harness makes a big

difference. A harness is a lot of the

reasons why you think coding with an LLM

has gotten so good. It's actually the

harness, the ability for the model to

use and abuse the system to gather

information and make smarter decisions.

A harness is going to be the thing that

gives the model access to tools in which

it can then go and make much more

informed decisions. So that means GLM

with no harness outperformed Opus 48 by

practically 10%, Jeopardy 55 by

practically 20%. And those both have

harnesses. And the thing about GLM that

you may not realize is that it's an

openweight model. I can go to hugging

face right now and download the model.

The thing is it's very very large. For

me to be able to run it with any level

of effectiveness, it could cost me about

$73 an hour. Now, if I was a bad actor,

you could see me actually doing that

because I could potentially make a lot

more than $73 an hour by scamming

people. And of course, it would be a

completely offline model for me to just

have. I wouldn't have to go and fork

over millions of dollars to be able to

have an onremise machine that could run

this. Instead, I could just run it from

the safety of my computer. Just doing a

quick little SSH up, bada bing, bada

boom, and just see the thing run. Not

only that, but Hermes agents, one of

these many agents, one of these many

harnesses that run models, has something

called a mixture of agents, meaning it

uses both 55 from Jippidy and also uses

four opus 48. And by combining those

both, it's doing better on benchmarks.

In other words, it's performing better.

I don't like honestly when someone says

it's 11% better, I have no idea what

that actually means cuz I don't know how

to quantify programming in that kind of

sense. I don't know how to be like it's

4% better. Like I don't know what that

means. But it's just saying, hey, it's

scoring higher on whatever these tests

are, which once again means that the

ability to do, you know, security

research and all this is actually kind

of generally available. the bugs that

Mythos has found, people are able to

find them. Also, with 48, you're going

to be able to find them with the mixture

of agents. GLM is going to be able to go

off and find these as well. As harnesses

improve, the ability to do this is going

to go way up. Microsoft also has a kind

of this like mixture of agents approach

in which they even score higher than

Mythos. So, it's all actually available.

It's just available to those with money.

If you have money and time and you can

set up your own harness, bada bing, bada

boom, you're going to have yourself a

hacking machine, which is ultimately

what Daario has been warning everybody

about. The scaling of of of open-source

models, I think it's going down a very

dangerous path. And if the if again, if

the path continues, I think we could get

to a very dangerous place. I think it's

worth saying some things on open source

models that are are clear to all the

experts but I want to make sure is is

understood by by this

>> clear to all the experts they uniformly

agree which is crazy because you know

classic experts I mean you can't even

get dentists to agree on the toothpaste

and he's over here getting experts to

agree to outcomes of stuff we haven't

measured yet kind of wild

>> committee which is when when you control

a model and you're deploying it you have

the ability to monitor it usage

>> it's really all you need to know you'd

like to monitor the usage and this is

the whole argument. This is the entire

safety argument which is just completely

broken. There's open source models in

which are doing absolutely excellent.

There's people that have money and

motivation to scam. They have access to

it. There's a whole bunch of people that

maintain a bunch of software that just

simply don't have access to these

models. They don't have access to these

open source frameworks cuz they're not

going to go spend, you know, $5,000 over

a weekend to maintain their little open-

source project that nobody actually

cares about in which the entire world

uses and everybody actually cares about.

They just don't realize they care about

it. And this is the whole problem of the

safety argument is that it's

fundamentally unsafe. The tools and the

harnesses are there. The mythos complete

hacking craze has already been

reproduced by plenty of other models and

people talking about it all the time on

the internet. There's already better

harnesses that are scoring higher,

whatever that means, on the same

benchmarks than Mythos. At the end of

the day, I think Michael, creator of

Effects TS, says it best. Going to need

to get something off my chest. The

discord on cyersc and mythos/ US

approach is stupid beyond imagination.

There should be no guardrails and the

most powerful models should be

immediately available to everyone. Want

to make sure the world is safer? Give

free usage to maintainers of core

infrastructure. If access is restricted,

be sure offensive teams will find their

way in. Somebody with insider access to

anthropic or open AI will be offered a

deal they can't refuse and leak the

model to malicious actors or a malicious

actor will end up developing an equally

capable model. Look at GLM. It's near

the same level of capability. Yeah,

maybe it's not mythos levels, but it's

not far off. Yes, not only the good

folks have money, but also the bad ones.

Exactly correct. The only thing you

achieve by restricting access is

unsafety

100%. Because that's the thing is right

now you could have an entire world that

exists in which people could be

preparing their piece of the world their

little corner of software making it more

secure if that's the true delivery of

these models to begin with. If that's

the case wouldn't you want the whole

world to walk in lock step? Why would

you want a select group of bad guys who

happen to have a couple million or the

ability to rent out enough Blackwell

Ultra GPUs from Nvidia to be able to

take advantage of everybody else while

we the plebs wait for the ability to

secure our own systems? Like the

fundamental premise of this all doesn't

make any sense. Like if Daario actually

wanted safety, he would go for

federation. It's the only way to make

the world safe is to have a lot of

people work on it at the same time. The

only way to make it unsafe is to allow

people to lag far behind what the bad

guys have access to. Ridiculous.

Absolutely ridiculous. And that's

because at the end of the day, anthropic

Daario over here, what they really want

is control. He says it right there

inside the little interview. If I can

just control the inputs, well, then I

can control the outputs. If I don't let

you say certain things, well, then you

can't say it to the model. I can protect

you from you. And ultimately, that's why

I call him Daddy Dario. He is the one

that's going to keep you safe from all

the bad guys or good guys, but you're

going to be safe because honestly, he

has your best interest in heart, which

is just control. I mean, sorry, safety.

Safety. He has your best interest in

heart. It's safety. That's what he

wants. Anyway, so there you go. That's

my whole opinion on safety in this

entire talk. I think it's all

ridiculous. I think this entire gated

access and everything is really not

doing anything good. I think at the end

of the day, the bad actors are still

going to be bad. The ability for Opus 48

to be able to hack at a near level as

mythos has already been cited a hundred

times all over these different studies.

It's not like we're actually really

having security. We're just having

security theater. And ultimately, it's a

play for control or it's a play for

Daario to be able to regulate his

opponents out of business. I'm not

actually really sure which one it is.

Probably both is my guess. Anyways, I

can't wait for Fable to come back out

again. That would be great, huh? I'd

love I'd love to have a little bit I'd

love to have a little bit of Fable.

Wouldn't you? Yeah, that'd be fantastic.

I' I'd use it again. Yeah. Fa fable.