HomeVideos

Be Careful w/ Skills

Now Playing

Be Careful w/ Skills

Transcript

261 segments

0:00

Sometimes in life you you wonder if you

0:02

should should laugh or if you should

0:04

cry. And this is this is one of those

0:07

points, okay? There are so many people

0:09

just absolutely getting owned right now

0:12

because of skills. Yes, people are

0:14

having skill issues. We're talking about

0:16

the highest order of skill issues. Now,

0:18

if you're not familiar with skills, uh

0:20

skills are effectively markdown files

0:22

that you can feed to your LLM while

0:24

you're making a prompt, and it should

0:25

give it extra context to be able to go

0:27

off and perform whatever you task you

0:29

want it to perform with a higher degree

0:31

of accuracy. You know, you have some

0:32

sort of obscure API, you give it all the

0:35

reasoning behind the API, it will just

0:36

use it better. It makes sense, right?

0:38

But what problems could happen with

0:40

skills? Like, what could actually go

0:41

wrong? Well, there's supply chain

0:43

issues. There's crypto bros. There's

0:44

skills that find skills, and you don't

0:46

know what you're going to get. skills

0:48

just hallucinate commands that don't

0:50

exist and you can even hide malicious

0:53

commands in HTML comments. So yes,

0:56

skills have a lot of issues. And so

0:58

let's kind of let's run down a few of

1:00

these. Now the first one right here,

1:01

which is probably one of my favorite X

1:03

articles I have ever uh read, which is

1:06

eating lobster souls part two, the

1:07

supply chain. Effectively, what happened

1:10

is this Jameson right here went on and

1:13

created a beautiful skill on Claude Hub.

1:15

Now remember, if you don't know what

1:17

Clawbot is, which was renamed to

1:18

Moldbot, which was renamed to Open Claw,

1:21

it's this like personal assistant that

1:23

you you give your all of your keys to

1:25

your life to, and it goes off and acts

1:27

on your behalf.

1:29

Nothing can go wrong there, don't you

1:31

worry. Well, it also happens to have a

1:33

skills hub so that if you wanted to say

1:35

interact with a new service, you can

1:36

give it the skills to pay those bills.

1:38

And so what Jameson did, of course, was

1:40

just simply create a fake skill, call it

1:43

something he knows that a bunch of bros

1:44

are going to get caught with, which was

1:46

what would Elon do, how Elon Musk would

1:49

break down problems. [laughter] I can't

1:52

believe that got people. Then abusing

1:54

one of the end points, he made that

1:56

package appear to be the most downloaded

1:58

and most popular of all the skills. But

2:01

luckily, he was nice and when it

2:02

executed, it was just like, "Dude, I

2:05

could have just like I could have owned

2:06

you. What are you doing? Why are you

2:09

First off, what is with this all? What

2:11

would Elon do? But second off, why would

2:14

you just execute random markdown files

2:16

you have not read? Do you not know the

2:19

danger of everything here? You gave it

2:22

the keys to your kingdom to Claudebot,

2:24

which has access to everything. My gosh,

2:27

I just cannot believe that people are

2:29

doing this. How did this happen? 2025

2:32

was the year of the human intervenor.

2:34

Everything should be ran through a

2:35

human. Hey, is that code? You review

2:37

that. Hey, is that a command? You review

2:39

that. You're going to whitelist the

2:41

commands you know are good. And you are

2:42

going to be diligent and vigilant about

2:45

the commands that aren't good because

2:47

you are the captain. Not in today's

2:49

world. You are no longer the captain.

2:51

You're just going to give it away for

2:53

free. What is 2026 deal? By the way, I'm

2:55

also again listening to Interstellar.

2:57

I'm just the song. It just I get so

2:59

pumped up right now. I'm just so pumped

3:02

up. But I think that honestly this is

3:03

one of the most clever kind of hacks

3:05

going on right here which is Zach right

3:07

here effectively did this. You can see

3:09

right here very very obvious and if you

3:11

go and you look at the raw GitHub it's

3:13

very very obvious it's right here. But

3:14

when you look at that skill on his page

3:18

it's not there. You don't see the

3:19

command because again markdown viewers

3:22

they're just so good at rendering

3:23

markdown aren't they? They'll even

3:25

render that HTML for you. Hey not a

3:27

problem. Oh is that a comment? Don't

3:29

worry we'll get rid of that. You don't

3:30

need to see that. Well, guess what? It's

3:32

hidden in plain sight. You could

3:35

actually have malicious commands that

3:37

you will never see because you looked at

3:40

a skill before downloading it on GitHub.

3:42

And this just goes to show like how

3:45

dangerous this world really is. There

3:47

there would be no world where I could

3:49

just like hand you a little executable

3:51

file and be like, "Hey, bro. Come on.

3:52

Just execute it for me." You know, I

3:53

just Hey, I you know, I just really need

3:54

it, okay? I'm a little lonely and if you

3:56

don't execute it, I'll be sad. You just

3:58

go, "No, I I don't want to do that." But

3:59

now, now for whatever reason, it's even

4:03

worse than kind of like the dangers of

4:04

npm and package managers in general. Now

4:07

it's like, hey, it's no longer you even

4:10

developing. It's just you handing

4:12

commands off to an LLM to run on your

4:14

behalf without ever even knowing what's

4:17

inside there. But I honestly think one

4:18

of the most kind of bizarre tangents

4:20

this entire world went on was this one

4:23

right here. So again, skills, they're

4:25

largely developed by LLMs. People that

4:27

are, you know, skill producers are

4:29

people that are just like, "Yo, LLM,

4:31

tell me how to use Cloudflare. No

4:33

mistakes." Boom. Cloudflare skill added

4:35

to my portfolio. Yo, I'm kind of like a

4:37

skills bro. You know, you got skill

4:39

issues. Don't you worry. I got your

4:41

back. Okay? I got some skills right here

4:43

in my coat. And so a lot of these what

4:45

ends up happening is that one of them

4:47

will make some sort of hallucination or

4:48

some sort of mistake and then it will

4:50

just spread through other skills. And

4:53

this is one of the best stories I have

4:56

ever read. Oh my gosh. Which is that

4:58

agents were spreading a hallucinated NPX

5:02

command because one of them hallucinated

5:04

it. Then other skills started using that

5:06

hallucination as a means to produce more

5:09

skills. And now at the time of writing

5:11

this which was days ago or by the way 10

5:13

days ago is that 237 skills on GitHub

5:17

had this imaginary command. And this

5:19

imaginary command is npx react code

5:21

shift which actually just sounds like

5:23

something that could exist and then boom

5:25

it would give it your source directory.

5:27

Now this would just simply fail. The

5:28

agent would move on. Most people didn't

5:31

even realize this was happening. So the

5:33

person who saw this all happening,

5:35

Charlie right here, he decided to

5:38

create, hey, what if I just create that

5:41

as something that is available on npm

5:45

and then when people try to execute this

5:47

fake command, it just goes to me. And

5:49

guess what it did? I don't even know

5:51

what to call this type of, you know,

5:53

vulnerability. Is this hallucination

5:55

squatting? Like what is this? This is

5:56

just a whole new level of insanity we

5:59

are reaching. But it it's somehow it's

6:01

still it just keeps getting worse. It

6:03

just keeps on happening. There there's

6:06

actually something called find skills

6:08

that Verscell's always telling you to

6:09

download. So get this one. Okay, this is

6:11

even better. Verscel you can just get a

6:13

skill on skillsh added. All you have to

6:16

do add the skill yourself and then

6:18

download it once and it appears in their

6:21

official listing. I had is even up there

6:24

like a decent amount at one point just

6:25

for fun. And it's not hard to get some

6:27

views, you know, acred. And so Verscel

6:30

always, you know, tells you, you know

6:31

what you should do? Hey, you're

6:32

downloading a skill. You should just let

6:35

us download any skill we think is good.

6:37

And what it does is it does a skill

6:39

called find skills. And find skills,

6:41

anytime you ask a question like how do I

6:43

do X, it goes and queries all the

6:47

available skills and it just will toss

6:49

that crap right onto your computer and

6:51

let it be executed. The best part is is

6:53

that those skills, they could be good

6:55

because remember skills, they're just a

6:57

they're just a pointer to GitHub pretty

6:58

much, which means that it could be a

7:00

good actor

7:02

until it's not a good actor. Somehow

7:04

we're just racing honestly towards like

7:06

the worst possible, least secure

7:08

ecosystem. I I genuinely don't

7:10

understand this. Like, do you understand

7:12

that just raw dogging text to an LLM

7:15

that has full permissions on your system

7:17

and potentially external systems is a

7:20

bad plan? Like do is this a hard concept

7:23

to understand? Well, it shouldn't be,

7:25

but somehow it is. And somehow we are

7:27

going full blast into it. Yes, crypto

7:30

bros are also joining in on the fun. A

7:32

lot of the skills are just a crypto

7:34

nightmare trying to get into all of your

7:36

businesses and take all of your

7:38

bitcoins. Get your fingers off my

7:40

Bitcoin, boy. I don't I don't even know

7:42

what that means. Anywh who, I can't tell

7:45

if this is like is this a public service

7:46

announcement or is this a rant? I don't

7:48

know. But it just I don't understand how

7:52

such massive ignorance has taken hold of

7:55

an entire ecosystem. Like one of the

7:57

parts that are kind of being sold with

7:58

AIS is that it raises the floor. Yeah.

8:01

Like I understand that people who can't

8:03

program can produce websites. Is that a

8:05

good thing? Is that a thing that

8:07

actually is really like what we need

8:09

more of? Well, I can say that there are

8:10

some good parts about it. My son getting

8:12

super into game development. He's just

8:14

vibing right now, but he's more and more

8:15

being like, I want to learn more about

8:17

Lua. I want to be able to program in

8:19

Roblox. And so I can see like the honest

8:21

the goodness of it all. Like I get that.

8:23

But also when the floor is so low and

8:26

it's being raised at such a rapid rate,

8:29

people don't even have the ability to

8:31

associate the potential risks of what

8:33

they're doing. And of course, at the

8:35

risk of sounding like a boomer, read the

8:37

skills. Open up Vim, navigate to the

8:40

skill and read it without some sort of

8:43

HTML viewer and just look at what you're

8:46

executing on your system before you use

8:48

it. I cannot believe I'm saying this in

8:51

a YouTube video. The name is the

8:53

Boomerin.

8:55

Nothing makes me feel more like a boomer

8:57

than telling people to read. Oh my Oh my

9:00

gosh. [music]

9:09

Hey, you're probably wondering why am I

9:12

in San Francisco? Well, I'm here for a

9:14

big event and I'm going to stream the

9:15

whole thing. It's going to be live on my

9:17

channel for the next 5 days. So, if

9:18

you're watching this video, it's

9:19

probably live right now.

Interactive Summary

The video discusses the emerging risks associated with 'skills', which are markdown files used to provide context to LLMs. While intended to improve accuracy, skills introduce significant security vulnerabilities. These include supply chain attacks where fake skills are created to appear popular, malicious commands hidden in HTML comments within skills, and the hallucination of commands by LLMs which then spread to other skills. A notable example is the "what would Elon do" skill that tricked users into downloading it, and the "npx react code shift" hallucination that led to 237 skills on GitHub referencing a non-existent command. The "find skills" feature by Vercel is also criticized for automatically downloading and executing potentially untrusted skills. The speaker expresses concern over the rapidly lowering bar for entry into development, leading to a lack of awareness about these risks, and urges users to read skills before executing them, comparing the situation to the dangers of npm package managers.

Suggested questions

7 ready-made prompts