WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=yWS0udrIOc8

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/yWS0udrIOc8.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.760 --> 00:00:13.040
All right.

s3
00:00:13.040 --> 00:00:14.720
Hey everybody.

s4
00:00:14.720 --> 00:00:17.160
Uh, yeah, yeah, yeah, yeah.

s5
00:00:17.160 --> 00:00:18.000
Okay.

s6
00:00:18.000 --> 00:00:19.480
I have about 18 minutes.

s7
00:00:19.480 --> 00:00:28.080
Um, I have a slide deck here that that Claude put together for me, but Claude still can't do very good slide decks, so I apologize for that.

s8
00:00:28.080 --> 00:00:30.800
Just just just focus on me, right?

s9
00:00:30.800 --> 00:00:32.960
Right here.

s10
00:00:33.120 --> 00:00:35.560
This is like the most fun thing to do every now.

s11
00:00:35.560 --> 00:00:41.240
If I if I start like running out of time, can can you like call time for me?

s12
00:00:41.240 --> 00:00:42.280
All right, perfect.

s13
00:00:42.280 --> 00:00:45.320
So, I'm Steve Yege.

s14
00:00:45.320 --> 00:00:48.800
I am here on behalf of Sneak today.

s15
00:00:48.800 --> 00:00:51.440
I am not getting paid for this talk or anything like that.

s16
00:00:51.440 --> 00:01:01.000
They were going to nice enough to buy me a ticket to the conference, which is cool, but I'm mostly here because I wanted to hang out with all of you, and I don't bite, and feel free to come and say hi afterwards.

s17
00:01:01.000 --> 00:01:04.280
I I wandered the halls amazed at all the brand logos.

s18
00:01:04.280 --> 00:01:16.840
Um, I'm up here because like the title of my talk is like "Agentic Security", but the real title of my talk is "Be Scared".

s19
00:01:18.440 --> 00:01:21.160
See, I Come on.

s20
00:01:21.160 --> 00:01:21.440
Come on.

s21
00:01:21.440 --> 00:01:22.200
Level with me.

s22
00:01:22.200 --> 00:01:24.760
Who here is scared?

s23
00:01:24.960 --> 00:01:27.160
That's good, but you're all in the security talks.

s24
00:01:27.160 --> 00:01:27.960
Of course you are.

s25
00:01:27.960 --> 00:01:31.000
The problem is they aren't out there.

s26
00:01:31.000 --> 00:01:32.720
Right?

s27
00:01:32.720 --> 00:01:41.880
I I went to Commonwealth I went to so many places last year, but I one of these big banks I was at, I was talking to their chief security architect in December,

s28
00:01:41.880 --> 00:01:49.760
and I was doing a Q&A, and talking about vibe coding, and everyone was like asking me questions, and I was you know, knocking them out of the park, and

s29
00:01:49.760 --> 00:02:04.840
he stands up real quiet at the end, and he goes, "Yeah, so he goes, "If everyone's shipping code at the same at sorry, at 10 times faster and the defect rate stays the same, the security defect, right?

s30
00:02:04.840 --> 00:02:08.160
The the vulnerability rate.

s31
00:02:10.000 --> 00:02:15.760
Then that doesn't that mean that the defects surface goes up by 10x?

s32
00:02:16.200 --> 00:02:18.280
And I it hit me so hard.

s33
00:02:18.280 --> 00:02:19.920
I I sank down to my knees.

s34
00:02:19.920 --> 00:02:23.200
I was just like what are we going to do about this?

s35
00:02:23.200 --> 00:02:26.360
Because it's a really really important point.

s36
00:02:26.360 --> 00:02:34.720
Cuz the subtle implied question is not if the defect rate stays the same, the defect rate's going to get worse.

s37
00:02:34.800 --> 00:02:39.720
A lot worse with AIs writing the code.

s38
00:02:39.720 --> 00:02:43.640
And so I didn't have an answer for him.

s39
00:02:43.640 --> 00:02:49.520
I actually have a partial answer for you today and I'll share it with you, but it's only a partial answer.

s40
00:02:49.520 --> 00:02:53.840
The real answer is you have to be scared of what's coming.

s41
00:02:53.840 --> 00:02:55.600
All right?

s42
00:02:55.600 --> 00:02:59.760
And it's becau it's bec it's it's it's Do I have a slide about this?

s43
00:02:59.760 --> 00:03:00.440
Yeah.

s44
00:03:00.440 --> 00:03:08.120
So it's not just that you're just like putting in more of the same defects cross-site scripting and blah blah blah.

s45
00:03:08.120 --> 00:03:09.680
You're still putting those in.

s46
00:03:09.680 --> 00:03:15.760
Fable wrote a an XSS vulnerability during the short time that I had with it.

s47
00:03:15.760 --> 00:03:16.480
Not its fault.

s48
00:03:16.480 --> 00:03:19.122
We'll talk about it in a minute.

s49
00:03:19.122 --> 00:03:19.200
[snorts]

s50
00:03:19.200 --> 00:03:22.160
But those are the old ones.

s51
00:03:22.160 --> 00:03:24.120
All right, we are we all know how to fix those.

s52
00:03:24.120 --> 00:03:31.959
There's new vulnerability types and new attack surfaces coming and they're here and many of them are incredibly well polished.

s53
00:03:31.959 --> 00:03:33.239
Like what's an example?

s54
00:03:33.239 --> 00:03:36.040
You guys know about slop squatting?

s55
00:03:36.040 --> 00:03:36.280
Right?

s56
00:03:36.280 --> 00:03:39.720
Where the AI hallucinates a package name.

s57
00:03:39.720 --> 00:03:43.720
Let's say that you want a graph database and so you're like, "Okay, I'm going to use this graph database."

s58
00:03:43.720 --> 00:03:45.440
And the AI goes, "Oh yeah, I know.

s59
00:03:45.440 --> 00:03:47.519
It's you know, it's graphy 123."

s60
00:03:47.519 --> 00:03:50.840
And it goes off to the package manager and downloads graphy 123.

s61
00:03:50.840 --> 00:04:00.120
And it builds and it runs, and the tests pass, and it looks right, but what it downloaded was a backdoor.

s62
00:04:00.120 --> 00:04:10.320
Because that graphy123 wasn't a real package, but somebody noticed that the LLMs are hallucinating its name, and they uploaded one that does exactly the same thing as the one it thought it was getting,

s63
00:04:10.320 --> 00:04:11.560
plus a vulnerability.

s64
00:04:11.560 --> 00:04:13.480
Bonus.

s65
00:04:13.480 --> 00:04:14.720
Yeah?

s66
00:04:14.720 --> 00:04:16.359
Is that scary?

s67
00:04:16.359 --> 00:04:18.160
Yeah, it should be.

s68
00:04:18.160 --> 00:04:19.120
It should be.

s69
00:04:19.120 --> 00:04:22.480
How do you even detect that that's happening?

s70
00:04:22.560 --> 00:04:41.360
So, so we're entering a world where everything you write every bit of code that you generate is going to have to get far more security scrutiny than it's ever had before.

s71
00:04:41.360 --> 00:04:41.960
Okay?

s72
00:04:41.960 --> 00:04:47.200
It's just you're not It's hard for me to convey how scared I am.

s73
00:04:47.200 --> 00:04:49.040
All right?

s74
00:04:49.040 --> 00:04:51.440
But let's start with where it gets generated.

s75
00:04:51.440 --> 00:04:54.440
Now, when I worked at Google, I worked really close with the TAP team.

s76
00:04:54.440 --> 00:04:55.720
They did tests.

s77
00:04:55.720 --> 00:04:58.080
The test automation platform, yeah?

s78
00:04:58.080 --> 00:05:02.040
And they uh So, they ran all of the unit tests and integration tests at Google.

s79
00:05:02.040 --> 00:05:07.280
They had a massive fleet, you know, and and they learned stuff about how bugs work.

s80
00:05:07.280 --> 00:05:15.360
And how bugs work is they have a lot they have like a life cycle where if you see the bug right away, you'll fix it.

s81
00:05:15.360 --> 00:05:22.480
And the longer the time goes for when you see a warning or some sort of issue, the longer that passes Okay?

s82
00:05:22.480 --> 00:05:28.880
It's got this sort of half-life of urgency, and all of a sudden, eh, it ain't really biting anyone anymore, okay?

s83
00:05:28.880 --> 00:05:30.720
This is how we handle all of our bugs.

s84
00:05:30.720 --> 00:05:43.360
At at Google, they recognized that this is such a human nature phenomenon that they worked really hard to to move the reporting of bugs as you were typing.

s85
00:05:43.480 --> 00:05:46.200
Because that's when you're most likely to fix it.

s86
00:05:46.200 --> 00:05:51.600
If you make a bug, and it goes, "By the way, there's a divide by zero here or there's a backdoor or a vulnerability or whatever."

s87
00:05:51.600 --> 00:05:53.200
You'll You'll fix it right there.

s88
00:05:53.200 --> 00:05:56.720
But, if it gets to code review time, you're like, "Is it really worth it?"

s89
00:05:56.720 --> 00:05:57.634
Right?

s90
00:05:57.634 --> 00:05:58.120
[snorts]

s91
00:05:58.120 --> 00:06:06.640
And the problem, folks, is that that works for all classes of bugs except for security.

s92
00:06:06.840 --> 00:06:08.800
There's no half-life on it biting you.

s93
00:06:08.800 --> 00:06:15.680
It's not like, "Oh, because users aren't getting bothered by this security vulnerability that it's not a problem over time."

s94
00:06:15.680 --> 00:06:18.760
The problem compounds over time.

s95
00:06:18.760 --> 00:06:19.800
Yeah?

s96
00:06:19.800 --> 00:06:31.280
So, you have to treat this class of vulnerabilities the way that Google treated their top vulnerabilities at the time, which is to surface them at the developer's fingertips.

s97
00:06:31.560 --> 00:06:34.600
What if the developer doesn't have any fingers?

s98
00:06:34.600 --> 00:06:36.934
I don't know how how many fingers the LLMs have.

s99
00:06:36.934 --> 00:06:37.280
[snorts]

s100
00:06:37.280 --> 00:06:44.520
Um then you need to surface it to the LLMs.

s101
00:06:44.520 --> 00:06:45.000
But, wait.

s102
00:06:45.000 --> 00:06:45.760
You say, "Wait.

s103
00:06:45.760 --> 00:06:46.000
Wait.

s104
00:06:46.000 --> 00:06:46.440
Wait.

s105
00:06:46.440 --> 00:06:46.760
Wait.

s106
00:06:46.760 --> 00:06:47.120
Wait.

s107
00:06:47.120 --> 00:06:47.400
Wait.

s108
00:06:47.400 --> 00:06:48.919
Fable's really smart."

s109
00:06:48.919 --> 00:06:52.400
Or at least it seemed that way for the 2 days I got to use it.

s110
00:06:52.400 --> 00:06:54.600
Uh can't Fable just write secure code?

s111
00:06:54.600 --> 00:06:55.880
Right?

s112
00:06:55.880 --> 00:06:57.320
I mean, come on.

s113
00:06:57.320 --> 00:06:57.960
Come on.

s114
00:06:57.960 --> 00:07:00.600
You You all know secure I mean, you're all here in this room, right?

s115
00:07:00.600 --> 00:07:03.240
You all know security is an arms race.

s116
00:07:03.240 --> 00:07:04.440
One that never ends.

s117
00:07:04.440 --> 00:07:06.800
One that's going up exponentially with Moore's law.

s118
00:07:06.800 --> 00:07:09.880
One that's going to get real uncomfortable when quantum comes along.

s119
00:07:09.880 --> 00:07:13.520
Thank goodness that's like 5 to 7 years away.

s120
00:07:13.600 --> 00:07:16.720
Uh my buddy says 15, so maybe somewhere in between.

s121
00:07:16.720 --> 00:07:18.360
But, in the meantime, right?

s122
00:07:18.360 --> 00:07:20.200
LLMs are a real problem.

s123
00:07:20.200 --> 00:07:21.720
Yeah?

s124
00:07:21.720 --> 00:07:25.160
So, how do you surface How do you surface the vulnerabilities?

s125
00:07:25.160 --> 00:07:27.400
Well, first I want to figure out how to do it myself.

s126
00:07:27.400 --> 00:07:30.040
I have a game I've been working on for 30 years.

s127
00:07:30.040 --> 00:07:34.840
I just had Fable do a security hardening pass during the time I had it.

s128
00:07:34.840 --> 00:07:42.160
Went through and it did all my cloud hardening, and it found a bunch of credentials, and and all a bunch of stuff.

s129
00:07:42.160 --> 00:07:42.435
And it

s130
00:07:42.435 --> 00:07:42.480
[snorts]

s131
00:07:42.480 --> 00:07:47.000
started to give me these vibes like, yep, yep, Harding pass looking pretty good.

s132
00:07:47.000 --> 00:07:48.880
So, I ran sneak, right?

s133
00:07:48.880 --> 00:07:50.800
And I I don't have the numbers here.

s134
00:07:50.800 --> 00:07:54.040
Yeah, I didn't I didn't include the numbers cuz I'm done.

s135
00:07:54.040 --> 00:07:58.840
But, um uh it found 241 vulnerabilities, right?

s136
00:07:58.840 --> 00:08:03.640
Just a ton that Fable hadn't even thought to look for, right?

s137
00:08:03.640 --> 00:08:09.480
Uh and it's because look, so I I I've told people about the rule of five.

s138
00:08:09.480 --> 00:08:17.640
When you do things with LLMs, often you have to get them to do up to four to five reviews of the work that they did before it's like actually ready to ship.

s139
00:08:17.640 --> 00:08:29.680
And it's because their cognitive process is very similar to ours and it goes through a draft and then a revision and then polish and editing until it's, you know, it's finally ready to go.

s140
00:08:29.680 --> 00:08:31.040
It's like painting a wall.

s141
00:08:31.040 --> 00:08:33.120
Some things you don't just do all at one pass.

s142
00:08:33.120 --> 00:08:36.360
You do them in multiple passes, right?

s143
00:08:36.360 --> 00:08:48.960
So, um security is one of the So, what I found I wrote a book on vibe coding last year.

s144
00:08:48.960 --> 00:08:53.680
I did did more vibe coding anything than anyone, you know, two two years ago.

s145
00:08:53.680 --> 00:09:02.640
And and what I found was um that they're really good at doing one thing at a time.

s146
00:09:02.640 --> 00:09:04.960
Even the really good models like Fable, right?

s147
00:09:04.960 --> 00:09:09.200
Just because of this multi-pass painting a wall phenomenon.

s148
00:09:09.200 --> 00:09:16.760
You got to give them one task at a time, which means you can't give them security at the same time as you give them correctness.

s149
00:09:16.760 --> 00:09:19.160
They'll do a half-ass job of both.

s150
00:09:19.160 --> 00:09:23.280
And you don't want a half-ass job of either of those, it turns out.

s151
00:09:23.280 --> 00:09:25.920
So, you do it in two passes, right?

s152
00:09:25.920 --> 00:09:25.940
Now,

s153
00:09:25.940 --> 00:09:26.160
[snorts]

s154
00:09:26.160 --> 00:09:33.800
it five It's been five months now, but five months ago I wrote an essay called Software Survival 3.0 where I talked about what software has to do to survive

s155
00:09:33.800 --> 00:09:35.960
when LLMs can synthesize it all.

s156
00:09:35.960 --> 00:09:45.560
I don't know if any of you all saw that, but the basic the basic gist of it is that LLMs can synthesize any software that they want, but they're very lazy

s157
00:09:45.560 --> 00:09:46.760
in a good way, right?

s158
00:09:46.760 --> 00:09:54.800
Lazy in like they don't want to spend tokens if they don't have to, because that's money and power and bad for the planet and bad for your wallet and so on, right?

s159
00:09:54.800 --> 00:10:00.400
And so they use tools to help them whenever it can save tokens, right?

s160
00:10:01.080 --> 00:10:09.520
So tying it all together, if the LLMs are doing the coding and they're happy to use tools to help them offload cognition, you see where this is going?

s161
00:10:09.520 --> 00:10:10.680
Give them sneak.

s162
00:10:10.680 --> 00:10:12.600
Give them chain guard.

s163
00:10:12.600 --> 00:10:19.200
And I still think there's a missing piece in this picture that I'll tell you about at the end.

s164
00:10:19.200 --> 00:10:22.240
Um chain I don't know if you all know about chain guard.

s165
00:10:22.240 --> 00:10:33.800
Chain guard uh Chain guard is a supply chain that you sign up for and they give you images that have been pre-vetted to not have vulnerabilities and they update them.

s166
00:10:33.800 --> 00:10:36.120
So it's your inputs, okay?

s167
00:10:36.120 --> 00:10:39.360
And then sneak handles everything else.

s168
00:10:39.360 --> 00:10:47.040
The code that you write, the code that the LLM writes, the dependencies that you're pulling in from slot squatting, right?

s169
00:10:47.040 --> 00:10:48.680
The innocent stuff.

s170
00:10:48.680 --> 00:10:57.200
It can find My understanding is that sneak can actually find vulnerabilities that are proprietary, that only they know about because they're ahead of the CVE registry.

s171
00:10:57.200 --> 00:10:59.480
I There's some truth to that.

s172
00:10:59.480 --> 00:11:08.720
When I ran them on my code base, it didn't find any vulnerabilities that weren't already public CVEs, but damn, it was easy to [clears throat] use, right?

s173
00:11:09.120 --> 00:11:17.200
I think that a tool like sneak is going to give your LLMs superpowers, okay?

s174
00:11:17.200 --> 00:11:24.280
Cuz what you do is you add it as a pass to the prompt that you give them for whatever they're doing.

s175
00:11:24.280 --> 00:11:37.200
And say one last thing to look at and have them run your security analysis, all of the tools, get the open source ones, get the sneak one, get the chain guard one, get the all of them and have them

s176
00:11:37.200 --> 00:11:38.920
check each other's work, too.

s177
00:11:38.920 --> 00:11:39.040
Right?

s178
00:11:39.040 --> 00:11:42.440
If you want to get really serious about this on lunch time, right?

s179
00:11:42.440 --> 00:11:46.800
But secure your supply chain because Five Eyes is warning us.

s180
00:11:46.800 --> 00:11:47.800
Yeah?

s181
00:11:47.800 --> 00:11:48.880
My poor my poor game.

s182
00:11:48.880 --> 00:11:53.120
By the way, please please I just told you that my game has 241 vulnerabilities.

s183
00:11:53.120 --> 00:11:54.320
Don't go hack my game.

s184
00:11:54.320 --> 00:11:55.880
Give me a couple days to fix the bugs.

s185
00:11:55.880 --> 00:11:56.560
Yeah?

s186
00:11:56.560 --> 00:11:57.640
We all good here?

s187
00:11:57.640 --> 00:11:58.320
Good.

s188
00:11:58.320 --> 00:12:02.473
All right, but if you do hack it, you're going to bother like five players, all right?

s189
00:12:02.473 --> 00:12:02.520
[snorts]

s190
00:12:02.520 --> 00:12:03.480
Okay.

s191
00:12:03.480 --> 00:12:04.840
They're very loyal.

s192
00:12:04.840 --> 00:12:14.960
Um look, Five Eyes, which is like a bunch of right governing it's it's big countries that are have have their eye on on the cybersecurity, you know, landscape.

s193
00:12:14.960 --> 00:12:18.720
They just announced that it is now months, not years, until it starts happening.

s194
00:12:18.720 --> 00:12:20.720
It.

s195
00:12:21.480 --> 00:12:23.920
You all know what it is, right?

s196
00:12:23.920 --> 00:12:29.840
It is when open source models catch up to Mythos.

s197
00:12:30.200 --> 00:12:34.960
Does anybody here believe open source models are going to catch up to Mythos?

s198
00:12:35.680 --> 00:12:38.200
Interesting that it's about 50/50.

s199
00:12:38.200 --> 00:12:41.280
Anyone got a time frame in mind?

s200
00:12:42.920 --> 00:12:43.839
Who said December?

s201
00:12:43.839 --> 00:12:46.280
That's pretty accurate.

s202
00:12:47.200 --> 00:12:48.520
Cheater.

s203
00:12:48.520 --> 00:12:49.920
Yeah, it's about 7 months.

s204
00:12:49.920 --> 00:12:52.920
So, actually it's shrinking, so it's probably about 6 months now.

s205
00:12:52.920 --> 00:12:54.000
Yeah.

s206
00:12:54.000 --> 00:12:58.320
And Mythos is real real good at hacking your systems.

s207
00:12:59.880 --> 00:13:10.800
And and just just remember you can't trust it to automatically write good code any more than you can trust it to write elegant code by default.

s208
00:13:10.800 --> 00:13:12.080
That's a separate concern.

s209
00:13:12.080 --> 00:13:13.920
It's a separate pass.

s210
00:13:13.920 --> 00:13:17.960
You can't expect it to write performant code by default.

s211
00:13:17.960 --> 00:13:19.320
That's another pass.

s212
00:13:19.320 --> 00:13:19.920
You see what I'm saying?

s213
00:13:19.920 --> 00:13:24.680
You can't expect it to necessarily write the code according to your company coding standards, okay?

s214
00:13:24.680 --> 00:13:31.960
These are all passes that go through your code and I just want you to remember that security should be your first one and your last one.

s215
00:13:31.960 --> 00:13:32.680
Okay?

s216
00:13:32.680 --> 00:13:34.760
Give it extra.

s217
00:13:34.760 --> 00:13:43.240
Okay, and the last thing I wanted to talk to you about first of all, go do all this, and second of all, the last thing is really true true truthfully,

s218
00:13:43.240 --> 00:13:44.200
okay?

s219
00:13:44.200 --> 00:13:46.080
Dial it in here, folks.

s220
00:13:46.080 --> 00:13:55.040
Go to your families offline, like in person, and get your your code words refreshed.

s221
00:13:55.040 --> 00:14:07.920
Cuz another kind of scam that's coming along is you get a call from a family member who's in distress, and they need money, and it's very convincing, and there's a video of them, and you're going to need a way to distinguish them from from AI.

s222
00:14:07.920 --> 00:14:09.280
Okay?

s223
00:14:09.280 --> 00:14:15.240
It's months away, and some of your families are going to be slow to catch on to this stuff, but bank accounts will be drained.

s224
00:14:15.240 --> 00:14:19.480
I heard that Congress was given secret demos of draining bank accounts.

s225
00:14:19.480 --> 00:14:23.520
I've been scared of this for close to 2 years.

s226
00:14:23.520 --> 00:14:34.320
I heard one talk from a security researcher almost 2 years ago at ETLS Las Vegas, and he stood up in front of the crowd, and he said, "You're all not scared enough of what's coming."

s227
00:14:34.320 --> 00:14:37.560
It'll affect you personally, not just your company.

s228
00:14:37.560 --> 00:14:38.360
Okay?

s229
00:14:38.360 --> 00:14:39.760
So, that's my message to you.

s230
00:14:39.760 --> 00:14:42.950
It's not a message of hope and positivity today,

s231
00:14:42.950 --> 00:14:43.680
[laughter]

s232
00:14:43.680 --> 00:14:53.640
but it is a it is a message that that's that should be clear crystal clear is that there are tools, open-source tools, free tools, commercial tools, okay?

s233
00:14:53.640 --> 00:14:56.400
Techniques, practices, okay?

s234
00:14:56.400 --> 00:15:01.400
That you can use right now to get started on fighting in this arms race and protecting yourself.

s235
00:15:01.400 --> 00:15:02.600
And that's all I've got today.

s236
00:15:02.600 --> 00:15:04.565
Thank you.

s237
00:15:04.565 --> 00:15:06.565
[applause]

s238
00:15:09.150 --> 00:15:11.150
[applause]

s239
00:15:13.560 --> 00:15:15.560
Questions?

s240
00:15:17.120 --> 00:15:19.440
We ready to go?

s241
00:15:30.600 --> 00:15:31.480
Sweet.

s242
00:15:31.480 --> 00:15:33.000
Um just super simple.

s243
00:15:33.000 --> 00:15:35.560
What has surprised you recently?

s244
00:15:35.560 --> 00:15:37.840
In the world of AI coding.

s245
00:15:37.840 --> 00:15:42.000
What has surprised me in the world of AI coding?

s246
00:15:42.080 --> 00:15:44.800
Well, I'm not really super representative.

s247
00:15:44.800 --> 00:15:49.840
I spend a lot of my time trying to predict the future by like hammering agents really, really hard.

s248
00:15:49.840 --> 00:15:51.040
Yeah.

s249
00:15:51.040 --> 00:16:02.000
Um so um Uh you know, one of the surprises and I shouldn't have been surprised, but one of the surprises is that AI is moving faster than the world is moving.

s250
00:16:02.000 --> 00:16:08.760
Uh tech is moving faster than society can move.

s251
00:16:08.760 --> 00:16:20.440
And the surprises show up when friends, smart friends, resist uh you know, the inevitability of AI and they call it psychosis or they right?

s252
00:16:20.440 --> 00:16:23.120
Or they poo-poo it and they say, "Well, it'll never be actually smart."

s253
00:16:23.120 --> 00:16:23.480
Or whatever.

s254
00:16:23.480 --> 00:16:25.160
They they can't see the curve.

s255
00:16:25.160 --> 00:16:26.400
Right?

s256
00:16:26.400 --> 00:16:28.839
And uh and that surprises me.

s257
00:16:28.839 --> 00:16:30.600
Uh maybe it shouldn't.

s258
00:16:30.600 --> 00:16:33.720
Um it's for it shows a sort of tunnel vision.

s259
00:16:33.720 --> 00:16:39.760
I think people have a tendency to look about 3 months back and about 3 months forward and be like, "Oh, it looks pretty flat."

s260
00:16:39.760 --> 00:16:40.280
Right?

s261
00:16:40.280 --> 00:16:49.720
But but and so that that surprises me that people aren't honestly that people aren't more scared and that and by the same by the flip side that people aren't more excited by it.

s262
00:16:49.720 --> 00:16:50.400
Right?

s263
00:16:50.400 --> 00:16:56.200
Cuz you know once you actually, you know, once once you get it, I mean you you don't even want to be here.

s264
00:16:56.200 --> 00:16:59.200
You How many of you are running Claude code right now?

s265
00:16:59.200 --> 00:17:00.600
Most of you, right?

s266
00:17:00.600 --> 00:17:01.760
It's really fun.

s267
00:17:01.760 --> 00:17:07.600
So, you know, I mean like that's a surprise, too, that the world is pushing back so hard on that, right?

s268
00:17:07.600 --> 00:17:09.640
We're in an awkward phase.

s269
00:17:09.640 --> 00:17:11.439
We'll get through it.

s270
00:17:11.439 --> 00:17:13.040
Any other questions?

s271
00:17:13.040 --> 00:17:13.920
Woah.

s272
00:17:13.920 --> 00:17:14.400
All right.

s273
00:17:14.400 --> 00:17:16.920
Well, you pick.

s274
00:17:19.640 --> 00:17:21.760
Feel free to bail also.

s275
00:17:21.760 --> 00:17:22.640
You don't have to stay.

s276
00:17:22.640 --> 00:17:24.079
I big fan.

s277
00:17:24.079 --> 00:17:28.160
Um, what's the like most impressive uh thing you've seen Gastown do?

s278
00:17:28.160 --> 00:17:30.960
And like how much human intervention was involved or steering?

s279
00:17:30.960 --> 00:17:32.200
Oh, Gastown?

s280
00:17:32.200 --> 00:17:33.240
Yeah.

s281
00:17:33.240 --> 00:17:34.160
Gastown.

s282
00:17:34.160 --> 00:17:36.880
Gastown was a lot of fun in January.

s283
00:17:36.880 --> 00:17:42.400
Um, yeah, uh Gastown is a Beads machine and I still use Beads.

s284
00:17:42.400 --> 00:17:48.800
And I'm working I want to I was talking to Angie, I want to donate Beads to the Agentyc Foundation, you know.

s285
00:17:48.800 --> 00:17:51.040
We're going to we're going to put multiple backends on it.

s286
00:17:51.040 --> 00:17:53.440
Beads is a task tracker, right?

s287
00:17:53.440 --> 00:17:56.600
Beads is how you do Boris Cherney loops.

s288
00:17:56.600 --> 00:17:59.480
You know how Boris is like you shouldn't be prompting your agent?

s289
00:17:59.480 --> 00:18:06.720
Has anybody here actually like successfully How often do you get Claude to actually run all night for you?

s290
00:18:06.720 --> 00:18:08.160
Like for real run all night.

s291
00:18:08.160 --> 00:18:10.080
A few of you, right?

s292
00:18:10.080 --> 00:18:16.720
Like this is the next frontier, I think, of actually getting agents to run like for a long long time unsupervised.

s293
00:18:16.720 --> 00:18:21.120
You can do it with Beads by queuing up enough enough work and having them claim and all that.

s294
00:18:21.120 --> 00:18:24.160
And there are some other some other systems that'll do that.

s295
00:18:24.160 --> 00:18:27.120
It was really fun when Gastown did this for me automatically once.

s296
00:18:27.120 --> 00:18:30.440
I filed a whole bunch of Beads and they disappeared and I was like, "Oh, no, another bug.

s297
00:18:30.440 --> 00:18:31.240
My Beads disappeared."

s298
00:18:31.240 --> 00:18:32.857
And what had actually happened was that

s299
00:18:32.857 --> 00:18:32.920
[snorts]

s300
00:18:32.920 --> 00:18:36.160
one of the agents just found them and just implemented everything, right?

s301
00:18:36.160 --> 00:18:39.120
So I was like, "Whoa, I really like swarms now."

s302
00:18:39.120 --> 00:18:40.360
Yeah.

s303
00:18:40.360 --> 00:18:41.880
Fun times.

s304
00:18:41.880 --> 00:18:47.600
Does anybody here regularly work with more than 10 coding agents at once?

s305
00:18:48.120 --> 00:18:49.520
You see like not many, right?

s306
00:18:49.520 --> 00:18:54.400
The world is still in the we're still kind of like prompting and using a few here and there, right?

s307
00:18:54.400 --> 00:18:58.360
It's going to accelerate really fast next year.

s308
00:18:58.360 --> 00:19:00.200
Other questions?

s309
00:19:00.200 --> 00:19:01.000
You.

s310
00:19:01.000 --> 00:19:01.880
You can just say it.

s311
00:19:01.880 --> 00:19:02.360
Yeah.

s312
00:19:02.360 --> 00:19:07.560
Um, your talk focused a lot on vulnerabilities from agents writing code.

s313
00:19:07.560 --> 00:19:12.160
I'm curious what are your thoughts about agents taking actions

s314
00:19:13.520 --> 00:19:16.640
Yeah, so that was the third dimension that I really wanted to talk about.

s315
00:19:16.640 --> 00:19:21.760
It's just I don't really have time, but like my friends over at Tessell, I'm advising them.

s316
00:19:21.760 --> 00:19:23.720
They're actually doing this, right?

s317
00:19:23.720 --> 00:19:29.080
There's just this whole space of who's looking up who's looking over your agent's shoulders.

s318
00:19:29.080 --> 00:19:30.880
I had this conversation just now.

s319
00:19:30.880 --> 00:19:36.000
Like I tell everyone, everyone's just starting to stand up agents like 24/7.

s320
00:19:36.040 --> 00:19:42.040
Like processing queues, responding to events, like agents that actually do stuff, right?

s321
00:19:42.040 --> 00:19:47.360
24/7 and I I I encourage people to think adversarially.

s322
00:19:47.360 --> 00:19:55.960
I I think of adversarial groups of agents tasked with doing that queue management cuz one agent will always eventually screw it up.

s323
00:19:55.960 --> 00:19:57.120
Right?

s324
00:19:57.120 --> 00:19:58.840
So, you got to have those supervisors.

s325
00:19:58.840 --> 00:20:01.280
And so, there's whole systems emerging here.

s326
00:20:01.280 --> 00:20:07.320
Kind of can go out and go look at all of your and say and start to like you know, do do do that hardening stuff.

s327
00:20:07.320 --> 00:20:10.360
Like do they really need all those credentials on that service account?

s328
00:20:10.360 --> 00:20:11.400
Really, right?

s329
00:20:11.400 --> 00:20:13.960
Only for this one action, maybe we can like separate this one out.

s330
00:20:13.960 --> 00:20:15.200
That kind of thing, right?

s331
00:20:15.200 --> 00:20:29.120
This is a brand new frontier, but it's one, ironically, that even though there's kind of an almost nothing out there, there's some experimental stuff, you still have to be thinking about it right now and designing a solution in-house right now, right?

s332
00:20:29.120 --> 00:20:37.600
Because otherwise your engineers are going to spin up or your non-engineers are going to spin up a bunch of agents with way too many permissions and then, right?

s333
00:20:37.600 --> 00:20:41.120
As soon as a bear munches into the igloo, everyone's dead, right?

s334
00:20:41.120 --> 00:20:43.080
That's the old security analogy.

s335
00:20:43.080 --> 00:20:45.169
I don't know if they still use that one anymore.

s336
00:20:45.169 --> 00:20:45.760
[clears throat]

s337
00:20:45.760 --> 00:20:46.800
Yes.

s338
00:20:46.800 --> 00:20:48.080
Hey Steve, nice to see you.

s339
00:20:48.080 --> 00:21:00.320
Um I'm curious in terms of like this the best practices that you've seen uh that you use personally or or maybe you haven't tried, but um particularly with regards to prompt injection.

s340
00:21:00.320 --> 00:21:00.960
So,

s341
00:21:00.960 --> 00:21:07.439
Yeah, so I was supposed to talk that during my right during during my speech here, I I I wanted to mention it.

s342
00:21:07.439 --> 00:21:07.560
[snorts]

s343
00:21:07.560 --> 00:21:14.080
There are a whole bunch of attacks happening on the training and on the prompting side, right?

s344
00:21:14.080 --> 00:21:15.954
So, on training and inference.

s345
00:21:15.954 --> 00:21:16.320
[snorts]

s346
00:21:16.320 --> 00:21:20.120
And so, the bad guys find ways to sneak in stuff, right?

s347
00:21:20.120 --> 00:21:33.800
Um and it's like like like the simplest version is the new XSRF where like the user puts in some some text and then some bad actor puts in some extra text saying disregard everything and do the following, right?

s348
00:21:33.800 --> 00:21:36.720
And then they just get more sophisticated from there.

s349
00:21:36.720 --> 00:21:42.920
Um I mean, I don't have any good answers for you other than like this is real.

s350
00:21:42.920 --> 00:21:44.960
It's kind of an education problem at this point.

s351
00:21:44.960 --> 00:21:47.000
You need to get everybody thinking about it, right?

s352
00:21:47.000 --> 00:21:56.960
And then you I feel like there are like new security roles about ready to emerge inside of companies, agentic security that it's kind of an extension of what they're already doing.

s353
00:21:56.960 --> 00:21:59.440
Who's already doing this?

s354
00:21:59.440 --> 00:21:59.800
Right?

s355
00:21:59.800 --> 00:22:07.200
Yeah, so you you've already got people that are going out and looking after the the sort of security of your agents that are in that are deployed in the

s356
00:22:07.200 --> 00:22:08.360
Yeah, so you're way ahead of everyone.

s357
00:22:08.360 --> 00:22:10.080
I'd love to come talk to you later.

s358
00:22:10.080 --> 00:22:11.400
Ha.

s359
00:22:11.400 --> 00:22:14.480
This is all brand new stuff, yeah?

s360
00:22:14.480 --> 00:22:16.480
Cool.

s361
00:22:29.315 --> 00:22:31.315
[music]