AI Native DevCon 2026 New York — practical agent skills grounded in conference talks
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Dan Lorenc · AI Native DevCon New York 2026
Recording: https://www.youtube.com/watch?v=2uuV0_9snGA
Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.
For topic navigation, see outline.md.
Cue 1 — 00:00:00,561 --> 00:00:04,421
Yes, I'm Dan Lorenc, co-founder and CEO of a company called Chainguard.
Cue 2 — 00:00:04,641 --> 00:00:08,141
We do open source software, supply chain security.
Cue 3 — 00:00:08,201 --> 00:00:12,241
And like I said, I want to talk about what starts to break as we start moving at
Cue 4 — 00:00:12,281 --> 00:00:16,401
machine speed. I'm not going to be talking about inside of a single organization,
Cue 5 — 00:00:16,441 --> 00:00:19,721
CI systems falling over, going down, that kind of thing.
Cue 6 — 00:00:19,761 --> 00:00:22,301
I'm going to be doing this more from an ecosystem perspective.
Cue 7 — 00:00:22,521 --> 00:00:25,521
We sit in the middle of a bunch of different supply chains at Chainguard.
Cue 8 — 00:00:25,561 --> 00:00:29,681
You all do as well the open source software that you build, the products that you
Cue 9 — 00:00:29,721 --> 00:00:34,001
ship to your customers as everyone starts moving faster and faster at different
Cue 10 — 00:00:34,041 --> 00:00:38,361
rates and different speeds, that friction in the supply chain starts to cause a lot
Cue 11 — 00:00:38,401 --> 00:00:41,141
of issues. So that's kind of what I'm going to zoom in on today.
Cue 12 — 00:00:42,161 --> 00:00:46,721
So in case you didn't notice, last month was Cybersecurity Awareness Month.
Cue 13 — 00:00:46,761 --> 00:00:53,161
And to celebrate, we got probably the roughest week for anybody paying attention to
Cue 14 — 00:00:53,201 --> 00:00:55,461
software vulnerabilities in recent memory.
Cue 15 — 00:00:56,401 --> 00:01:00,181
On one day of that week, the week of I think it was the 19th,
Cue 16 — 00:01:00,281 --> 00:01:04,481
there were at least five of some of the most severe software vulnerabilities in open
Cue 17 — 00:01:04,521 --> 00:01:07,561
source software reported, all in one day.
Cue 18 — 00:01:07,601 --> 00:01:11,441
There have been worse ones, but in any normal year, any one of these would have been
Cue 19 — 00:01:11,481 --> 00:01:14,181
a headline vulnerability of the year candidate.
Cue 20 — 00:01:14,281 --> 00:01:15,961
And there were five all in one day.
Cue 21 — 00:01:16,001 --> 00:01:18,901
The week after, there was another massive one in the Linux kernel.
Cue 22 — 00:01:19,001 --> 00:01:20,901
This is kind of the new normal.
Cue 23 — 00:01:20,961 --> 00:01:23,201
I don't know why everyone picked Cybersecurity Awareness Month,
Cue 24 — 00:01:23,241 --> 00:01:28,041
but they did for this. But this is kind of a taste for the industry of what's what's
Cue 25 — 00:01:28,081 --> 00:01:30,121
been happening inside of the frontier labs.
Cue 26 — 00:01:30,161 --> 00:01:34,881
And the folks with these vulnerability research teams that are pointing these models
Cue 27 — 00:01:34,921 --> 00:01:40,001
at open source software. The main interesting point, and kind of the scary point for
Cue 28 — 00:01:40,041 --> 00:01:43,281
all of this, is that all of these vulnerabilities, the one that's found in PHP,
Cue 29 — 00:01:43,321 --> 00:01:44,981
the one in the compression algorithm.
Cue 30 — 00:01:46,041 --> 00:01:50,201
The one in GitLab, which was actually vulnerability down in Ruby's regular
Cue 31 — 00:01:50,241 --> 00:01:53,321
expression engine that was chained up through GitLab, the product.
Cue 32 — 00:01:53,361 --> 00:01:56,341
They had all been discovered months ago.
Cue 33 — 00:01:56,521 --> 00:01:59,721
Some of them even fixed months ago, but no one realized they were software
Cue 34 — 00:01:59,761 --> 00:02:04,721
vulnerabilities. I think the most recent discovery here was actually the PHP one,
Cue 35 — 00:02:04,761 --> 00:02:08,241
which was discovered simultaneously by a couple different research groups three
Cue 36 — 00:02:08,281 --> 00:02:12,781
months ago. The GitLab one had been public on Ruby's branch since April,
Cue 37 — 00:02:13,121 --> 00:02:17,961
the one similar people knew about these for a long time before anybody realized that
Cue 38 — 00:02:18,001 --> 00:02:21,701
there were vulnerabilities and before the CVEs got filed for these.
Cue 39 — 00:02:21,961 --> 00:02:26,161
And so if you take a look at this from the defender's advantage that Anthropic talks
Cue 40 — 00:02:26,201 --> 00:02:30,121
about as they launched the Glasswing program and stuff, and we need to give people
Cue 41 — 00:02:30,161 --> 00:02:33,761
that are trying to patch these libraries a head start before attackers get these
Cue 42 — 00:02:33,801 --> 00:02:35,861
same capabilities in their models.
Cue 43 — 00:02:36,081 --> 00:02:38,001
These were all found three plus months ago.
Cue 44 — 00:02:38,041 --> 00:02:42,521
And if the open models are behind the frontier models by some curve,
Cue 45 — 00:02:42,561 --> 00:02:45,561
we're already starting to kind of run out of time here for the defenders to actually
Cue 46 — 00:02:45,601 --> 00:02:49,641
get open source patched before people can start finding these and using them to
Cue 47 — 00:02:49,681 --> 00:02:50,941
attack companies.
Cue 48 — 00:02:52,801 --> 00:02:55,801
And so I like to use this diagram here to kind of talk through the pace of
Cue 49 — 00:02:55,841 --> 00:02:59,281
vulnerability management and where everyone sits in a supply chain historically.
Cue 50 — 00:02:59,321 --> 00:03:01,241
Right. Everyone sits at some point in this.
Cue 51 — 00:03:01,281 --> 00:03:04,041
It's a bunch of different buckets that are all kind of filling up.
Cue 52 — 00:03:04,081 --> 00:03:06,961
And we've kind of had equilibrium here for the last couple of decades.
Cue 53 — 00:03:07,001 --> 00:03:08,601
People are finding vulnerabilities.
Cue 54 — 00:03:08,641 --> 00:03:11,881
They get them patched, the patches get released.
Cue 55 — 00:03:11,921 --> 00:03:17,321
Security teams everywhere grab those patches, rebuild or redeploy your software
Cue 56 — 00:03:17,361 --> 00:03:20,081
systems. All of these have kind of been flowing at the same rate.
Cue 57 — 00:03:20,121 --> 00:03:23,941
Some software is harder to patch than others depending on how you deploy.
Cue 58 — 00:03:24,201 --> 00:03:27,001
Some is easier. Sometimes some projects fall behind.
Cue 59 — 00:03:27,041 --> 00:03:31,041
If the maintainers are overworked or sometimes the projects even get abandoned.
Cue 60 — 00:03:31,081 --> 00:03:34,501
But when you zoom out and take an ecosystem specific view of it,
Cue 61 — 00:03:34,641 --> 00:03:36,581
things kind of flow through like this.
Cue 62 — 00:03:37,161 --> 00:03:41,161
But unfortunately, as different parts of this start speeding up at different rates.
Cue 63 — 00:03:41,201 --> 00:03:44,801
You kind of get this the future is here, but it's not evenly distributed yet.
Cue 64 — 00:03:44,841 --> 00:03:49,121
Effect where we're starting to see some bottlenecks and some of these wells filling
Cue 65 — 00:03:49,161 --> 00:03:54,001
up a lot faster than others. And this is sort of where we are today.
Cue 66 — 00:03:54,041 --> 00:03:56,821
If you zoom in here, some of these things start to overflow.
Cue 67 — 00:03:57,201 --> 00:04:02,721
Anthropic on the Glasswing public disclosure site is talking about thousands and
Cue 68 — 00:04:02,761 --> 00:04:05,201
thousands that they've found. And you can see that they've actually stopped looking.
Cue 69 — 00:04:05,241 --> 00:04:08,081
They stopped trying to find more because they realized the bottleneck was actually
Cue 70 — 00:04:08,121 --> 00:04:13,101
getting this stuff fixed. And now it's not actually hard to get this stuff fixed.
Cue 71 — 00:04:13,281 --> 00:04:16,881
From a technical sense, these same agents that can find these vulnerabilities can
Cue 72 — 00:04:16,921 --> 00:04:19,421
generate pretty good patches for these vulnerabilities.
Cue 73 — 00:04:19,641 --> 00:04:20,841
But these are open source projects.
Cue 74 — 00:04:20,881 --> 00:04:23,921
You having a patch of it on your fork locally doesn't actually help you,
Cue 75 — 00:04:23,961 --> 00:04:26,981
because you have to get this out and release to the world.
Cue 76 — 00:04:27,241 --> 00:04:29,721
Maintainers might have a different take on how to patch some of this.
Cue 77 — 00:04:29,761 --> 00:04:31,401
There's different trade offs involved.
Cue 78 — 00:04:31,441 --> 00:04:33,041
Sometimes you can break compatibility.
Cue 79 — 00:04:33,081 --> 00:04:36,261
Sometimes you have to do it a different way to operate compatibility.
Cue 80 — 00:04:36,321 --> 00:04:38,001
So it's not really generating the patch.
Cue 81 — 00:04:38,041 --> 00:04:42,001
That's the issue. It's a patch that the community and the open source projects
Cue 82 — 00:04:42,041 --> 00:04:45,761
accept. That takes a while. Sometimes you can't even find the maintainers.
Cue 83 — 00:04:45,801 --> 00:04:49,441
Sometimes the maintainers inboxes are so full from all of the slop reports that
Cue 84 — 00:04:49,481 --> 00:04:52,161
they've been getting over the last couple of years, that they don't bother to read
Cue 85 — 00:04:52,201 --> 00:04:54,221
the real ones that are kind of hitting us right now.
Cue 86 — 00:04:55,321 --> 00:04:59,681
And so that first backlog a couple of months ago, I would have said that that first,
Cue 87 — 00:04:59,721 --> 00:05:03,221
well, on the left was the one that was about to overflow, but but it overflowed.
Cue 88 — 00:05:03,321 --> 00:05:07,601
We figured out how to get these things reported, and even the ones that projects
Cue 89 — 00:05:07,641 --> 00:05:11,321
haven't been able to acknowledge in time, the vulnerability disclosure window is
Cue 90 — 00:05:11,361 --> 00:05:14,701
starting to run out on those. And so they're going to start making their way public.
Cue 91 — 00:05:14,921 --> 00:05:16,801
You're going to start seeing more and more of these getting released,
Cue 92 — 00:05:16,841 --> 00:05:18,741
kind of like what happened a couple weeks ago.
Cue 93 — 00:05:19,121 --> 00:05:20,641
That's going to be the new normal.
Cue 94 — 00:05:20,681 --> 00:05:25,881
And we're going to start seeing that level of vulnerability release almost every
Cue 95 — 00:05:25,921 --> 00:05:27,441
single week for the foreseeable future.
Cue 96 — 00:05:27,481 --> 00:05:30,221
We just have to because these walls are going to overflow.
Cue 97 — 00:05:32,201 --> 00:05:34,601
This third one is now the one that has to catch up.
Cue 98 — 00:05:34,641 --> 00:05:37,701
And that's kind of where everyone that builds software sits.
Cue 99 — 00:05:38,961 --> 00:05:42,721
This is going to be a pretty violent catch up, because it's not just adjusting to
Cue 100 — 00:05:42,761 --> 00:05:44,521
the new state of vulnerability discovery.
Cue 101 — 00:05:44,561 --> 00:05:49,401
It's catching up to 3 to 6 months of accelerated vulnerability discovery that just
Cue 102 — 00:05:49,441 --> 00:05:51,141
hasn't reached everyone yet.
Cue 103 — 00:05:52,401 --> 00:05:54,801
People here, this is AI Native DevCon.
Cue 104 — 00:05:54,841 --> 00:05:56,521
People here probably ship pretty quickly.
Cue 105 — 00:05:56,561 --> 00:05:59,761
I've got agents running right now that are probably shipping to production while I'm
Cue 106 — 00:05:59,801 --> 00:06:03,401
tabbed away from the terminal, but not all software can ship that fast,
Cue 107 — 00:06:03,441 --> 00:06:07,721
critical infrastructure. You know, there's software running on the Mars rover.
Cue 108 — 00:06:07,761 --> 00:06:11,861
You can't really do continuous integration on the Mars rover in outer space.
Cue 109 — 00:06:12,321 --> 00:06:14,481
The grocery store attacks all over Europe.
Cue 110 — 00:06:14,521 --> 00:06:19,061
You can update firmware in a refrigerator, but you can't do that 50 times a day.
Cue 111 — 00:06:19,481 --> 00:06:22,881
These things are all connected to the internet, and as parts of the industry start
Cue 112 — 00:06:22,921 --> 00:06:24,921
moving faster and discovering more vulnerabilities.
Cue 113 — 00:06:24,961 --> 00:06:26,921
Some folks are going to be able to keep up and going to be able to ship,
Cue 114 — 00:06:26,961 --> 00:06:29,601
and they're going to be able to take these patches as fast as they need to.
Cue 115 — 00:06:29,641 --> 00:06:32,521
But other parts are really going to struggle, and that's where things start to get
Cue 116 — 00:06:32,561 --> 00:06:34,941
challenging from an ecosystem perspective.
Cue 117 — 00:06:37,521 --> 00:06:39,321
And so this is what's about to happen here, right.
Cue 118 — 00:06:39,361 --> 00:06:42,561
These dams don't leak. They're going to start to break.
Cue 119 — 00:06:42,601 --> 00:06:46,281
We can't keep finding and sitting on these vulnerabilities inside of all of the
Cue 120 — 00:06:46,321 --> 00:06:49,261
frontier labs forever. They have to make their way public.
Cue 121 — 00:06:49,881 --> 00:06:52,881
We're also starting to run out of time on the other end as attackers get these same
Cue 122 — 00:06:52,921 --> 00:06:55,601
models. They're getting independently discovered.
Cue 123 — 00:06:55,641 --> 00:06:58,681
And a lot of these fixes are also sitting in plain sight.
Cue 124 — 00:06:58,721 --> 00:07:02,441
People just don't realize how bad they are until somebody puts a full exploit chain
Cue 125 — 00:07:02,481 --> 00:07:03,301
together.
Cue 126 — 00:07:05,561 --> 00:07:08,241
So this is kind of the final graph here of the supply chain itself.
Cue 127 — 00:07:08,281 --> 00:07:12,321
Everyone here sits somewhere in this and you don't really get to pick your spot.
Cue 128 — 00:07:12,361 --> 00:07:16,121
If you ship software, you're somewhere here in the middle because you're depending
Cue 129 — 00:07:16,161 --> 00:07:20,321
on other software to ship it. People are finding vulnerabilities in the software.
Cue 130 — 00:07:20,361 --> 00:07:21,881
They've always been there. They're not injecting them.
Cue 131 — 00:07:21,921 --> 00:07:25,081
It's not their fault, but they're finding things that have been around for a decade.
Cue 132 — 00:07:25,121 --> 00:07:26,881
They're reporting them, things are getting released.
Cue 133 — 00:07:26,921 --> 00:07:28,561
Other distributions have to pick them up.
Cue 134 — 00:07:28,601 --> 00:07:32,741
Vendors package that into their software and you have to update all of this.
Cue 135 — 00:07:33,001 --> 00:07:37,441
I call this the Human Centipede of vulnerability disclosure.
Cue 136 — 00:07:37,481 --> 00:07:41,141
Everyone here is kind of sewn into it in a depressing way.
Cue 137 — 00:07:41,441 --> 00:07:42,721
You don't get to get out of it.
Cue 138 — 00:07:42,761 --> 00:07:47,481
And if you get a nice break for a month or two, you're going to regret that later as
Cue 139 — 00:07:47,521 --> 00:07:49,341
the backlog starts to flow.
Cue 140 — 00:07:52,921 --> 00:07:56,541
So breaking this down a little bit more into a nice two by two.
Cue 141 — 00:07:57,161 --> 00:08:00,081
There's a lot of stuff that has to change as an industry, and there's a lot of stuff
Cue 142 — 00:08:00,121 --> 00:08:02,521
that has to change in the way that we ship software.
Cue 143 — 00:08:02,561 --> 00:08:05,281
So we can kind of look at this from a macro and a micro perspective,
Cue 144 — 00:08:05,321 --> 00:08:08,521
what you can do to prepare yourself, but then also what we have to do as an industry
Cue 145 — 00:08:08,561 --> 00:08:12,281
to get off of this treadmill. And then there's a couple of lenses here to both from
Cue 146 — 00:08:12,321 --> 00:08:17,481
security and development. So we're going to start here with the macro view.
Cue 147 — 00:08:17,521 --> 00:08:24,441
As an industry, one of the scariest and most surprising pieces of all of this is
Cue 148 — 00:08:24,481 --> 00:08:27,921
that a lot of the vulnerabilities and a lot of the attacks and a lot of the chaining
Cue 149 — 00:08:27,961 --> 00:08:32,301
that's happening with the frontier models when you zoom in is on invisible
Cue 150 — 00:08:32,841 --> 00:08:38,041
vulnerabilities. These are things where projects just refactored stuff away or fix
Cue 151 — 00:08:38,081 --> 00:08:41,601
something that they thought was a bug, but it turns out that that code was actually
Cue 152 — 00:08:41,641 --> 00:08:45,401
vulnerable. They fixed that vulnerability, but nobody downstream knew that was a
Cue 153 — 00:08:45,441 --> 00:08:48,061
vulnerability, so they didn't know they had to upgrade.
Cue 154 — 00:08:48,481 --> 00:08:51,668
A famous example of this is actually kind of like Log4Shell.
Cue 155 — 00:08:51,708 --> 00:08:57,221
Log4j did not go through a nice private coordinated disclosure process or anything.
Cue 156 — 00:08:57,441 --> 00:09:00,841
Somebody just pushed a pull request up to fix what they thought was a bug in that
Cue 157 — 00:09:00,881 --> 00:09:04,641
repository. A couple hours later, one of the maintainers realized that that was not
Cue 158 — 00:09:04,681 --> 00:09:08,161
just a bug. That was a terrible vulnerability that had just been disclosed by
Cue 159 — 00:09:08,201 --> 00:09:10,561
accident, and everyone had to race to fix it.
Cue 160 — 00:09:10,601 --> 00:09:12,741
This kind of thing is happening every single day.
Cue 161 — 00:09:12,881 --> 00:09:17,201
The Linux kernel took a different approach, where for every single bug fix in the
Cue 162 — 00:09:17,241 --> 00:09:20,721
Linux kernel, they file a CVE because they don't know which ones are vulnerabilities
Cue 163 — 00:09:20,761 --> 00:09:24,001
or not. So they just call everything a vulnerability because it kind of is possibly
Cue 164 — 00:09:24,041 --> 00:09:27,721
a vulnerability down at that level, depending on how the kernel is configured and
Cue 165 — 00:09:27,761 --> 00:09:30,881
what you're running on it. The most applications don't do this.
Cue 166 — 00:09:30,921 --> 00:09:33,721
They just merge fixes, they don't really pay attention, and they only just those
Cue 167 — 00:09:33,761 --> 00:09:35,581
things that they actually know about.
Cue 168 — 00:09:36,641 --> 00:09:42,201
So if you look at a couple of the examples from last month, the 10.0 full remote
Cue 169 — 00:09:42,241 --> 00:09:48,701
code execution vulnerability in GitLab that was fixed upstream in Ruby,
Cue 170 — 00:09:48,761 --> 00:09:53,261
6 or 8 months prior to anybody noticing that that caused a vulnerability in GitLab
Cue 171 — 00:09:53,641 --> 00:09:56,561
because nobody knew that that fixed it, never actually got shipped into any
Cue 172 — 00:09:56,601 --> 00:09:59,641
downstream Ruby release, and GitLab never pulled up the fix.
Cue 173 — 00:09:59,681 --> 00:10:02,141
They would have done that six months ago if they had known.
Cue 174 — 00:10:02,841 --> 00:10:07,601
If you just take any kind of commodity, open model, even a bash script,
Cue 175 — 00:10:07,641 --> 00:10:11,121
you can look around at project commit history and find things that are likely
Cue 176 — 00:10:11,161 --> 00:10:15,481
vulnerabilities that can be chained into larger vulnerability chains without anybody
Cue 177 — 00:10:15,521 --> 00:10:17,941
even realizing that they fixed a security bug.
Cue 178 — 00:10:18,601 --> 00:10:22,761
And so if you're sitting downstream of this, waiting for patches to come to,
Cue 179 — 00:10:22,801 --> 00:10:25,441
or waiting for vulnerabilities to get filed before you upgrade,
Cue 180 — 00:10:25,481 --> 00:10:27,461
you're just going to fall even further behind.
Cue 181 — 00:10:27,561 --> 00:10:31,461
Most of the exploitable vulnerabilities never have a CVE filed for them.
Cue 193 — 00:11:14,014 --> 00:11:17,454
And the second big one from an industry perspective, the more you dig in here and
Cue 194 — 00:11:17,494 --> 00:11:23,254
the more you run Mythos or GPT Cyber or any of these, you realize that some of this
Cue 195 — 00:11:23,294 --> 00:11:26,154
old software that we're on is just unpatched.
Cue 196 — 00:11:26,414 --> 00:11:29,694
You can run one of these frontier models, find ten vulnerabilities,
Cue 197 — 00:11:29,734 --> 00:11:33,594
patch those ten vulnerabilities, run it again and find 20 more.
Cue 198 — 00:11:33,894 --> 00:11:38,414
There are just too many unsafe design patterns that are still out there in all of
Cue 199 — 00:11:38,454 --> 00:11:40,194
the software we're running today.
Cue 200 — 00:11:40,454 --> 00:11:43,974
Or if you get it perfectly right under perfect conditions, maybe you can keep
Cue 201 — 00:11:44,014 --> 00:11:47,354
vulnerabilities out, but most people can't do that.
Cue 202 — 00:11:47,494 --> 00:11:50,654
And we have to start thinking as an industry on how to get rid of these patterns.
Cue 203 — 00:11:50,694 --> 00:11:54,314
This is something that we've done dozens of times as a software industry.
Cue 204 — 00:11:55,014 --> 00:12:01,814
Back in the mid to late 2000s, early 2010s, Cross-site scripting was everywhere.
Cue 205 — 00:12:01,854 --> 00:12:06,694
If you click the wrong link in a Facebook DM, JavaScript could export your entire
Cue 206 — 00:12:06,734 --> 00:12:09,614
friend list and spam everyone on your timeline, right?
Cue 207 — 00:12:09,654 --> 00:12:11,374
These were incredibly hard to get, right?
Cue 208 — 00:12:11,414 --> 00:12:14,614
If you wrote any JavaScript, you were probably introducing this by accident.
Cue 209 — 00:12:14,654 --> 00:12:17,014
New frameworks, new technologies, and the browser came out.
Cue 210 — 00:12:17,054 --> 00:12:18,574
And now these are pretty much a thing of the past.
Cue 211 — 00:12:18,614 --> 00:12:21,594
You don't have to worry about these the same way as you write code today.
Cue 212 — 00:12:22,224 --> 00:12:27,894
The same thing for TLS on the internet outside of the certificate incident of that
Cue 213 — 00:12:27,934 --> 00:12:31,254
same week. With all of these vulnerabilities, you can mostly rely on the green
Cue 214 — 00:12:31,294 --> 00:12:33,454
checkbox being set across the browser.
Cue 215 — 00:12:33,494 --> 00:12:36,694
You don't have to worry about going to a Starbucks or a conference and logging onto
Cue 216 — 00:12:36,734 --> 00:12:40,134
the Wi-Fi here and somebody sniffing your password in plaintext.
Cue 217 — 00:12:40,174 --> 00:12:43,534
And we've invented new technologies time and time again to get rid of entire classes
Cue 218 — 00:12:43,574 --> 00:12:47,114
of vulnerabilities. That's what we have to continue doing here.
Cue 219 — 00:12:47,494 --> 00:12:50,334
There's the meme of rewrite it in Rust, all of this stuff.
Cue 220 — 00:12:50,374 --> 00:12:52,674
But memory safety issues are everywhere.
Cue 221 — 00:12:53,094 --> 00:12:57,534
I can count on one hand the number of C or C++ programs that we've looked at or
Cue 222 — 00:12:57,574 --> 00:13:01,294
scanned with our harness that did not actually have any remote code execution or
Cue 223 — 00:13:01,334 --> 00:13:03,774
similar bugs inside of it. It's possible.
Cue 224 — 00:13:03,814 --> 00:13:08,494
It's just so hard that almost no, no one without superhuman programming abilities
Cue 225 — 00:13:08,534 --> 00:13:09,554
can do this.
Cue 226 — 00:13:10,814 --> 00:13:14,574
Multifactor authentication. All of this phishing attacks basically don't exist
Cue 227 — 00:13:14,614 --> 00:13:16,514
anymore with these technologies.
Cue 228 — 00:13:17,014 --> 00:13:21,374
We need to start doing this across all of the code and all of the libraries that
Cue 229 — 00:13:21,414 --> 00:13:24,714
we're using. It's broader than just rewrite it in Rust, though.
Cue 230 — 00:13:24,934 --> 00:13:28,854
There's Java template libraries that are basically just dynamic code execution as a
Cue 231 — 00:13:28,894 --> 00:13:35,747
service. The Python YAML where pickling is basically just running
Cue 232 — 00:13:35,787 --> 00:13:40,774
unsafe code. We have to look forward, look forward and think about new technologies
Cue 233 — 00:13:40,814 --> 00:13:43,774
that can get rid of these. So we're not just patching everything every single day as
Cue 234 — 00:13:43,814 --> 00:13:45,154
they keep getting found.
Cue 235 — 00:13:48,094 --> 00:13:50,554
The good news here though, this isn't all scary.
Cue 236 — 00:13:50,894 --> 00:13:53,714
This stuff is actually kind of fun if you get it right.
Cue 237 — 00:13:54,214 --> 00:13:57,734
These same AI technologies that are causing us to all panic right now,
Cue 238 — 00:13:57,774 --> 00:14:00,554
and finding vulnerabilities faster than we can fix them.
Cue 239 — 00:14:00,614 --> 00:14:02,654
These same things can be used to upgrade our software.
Cue 240 — 00:14:02,694 --> 00:14:06,814
They can be used to modernize all of these tech stacks, and not just inside of
Cue 241 — 00:14:06,854 --> 00:14:09,814
startups, inside of some of the largest organizations, and the most critical
Cue 242 — 00:14:09,854 --> 00:14:12,014
infrastructure in the world. It's the farthest behind.
Cue 243 — 00:14:12,054 --> 00:14:13,794
It has the highest benefit there.
Cue 244 — 00:14:14,094 --> 00:14:16,694
I think AI is probably the best business model ever invented,
Cue 245 — 00:14:16,734 --> 00:14:20,514
because it's the cause of and solution to all of our problems at this point.
Cue 246 — 00:14:21,974 --> 00:14:24,414
What's slowing all of this down right now is just a lack of focus,
Cue 247 — 00:14:24,454 --> 00:14:27,334
and a lack of a destination to get to as part of as part of all of this
Cue 248 — 00:14:27,374 --> 00:14:29,474
modernization and all of these upgrades.
Cue 249 — 00:14:32,834 --> 00:14:35,994
OpenAI, a bunch of the labs have started these patch the planet initiatives where
Cue 250 — 00:14:36,034 --> 00:14:39,394
they're funding maintainers to give away free tokens, free agent tools,
Cue 251 — 00:14:39,434 --> 00:14:42,474
and everything to start patching these vulnerabilities, which is half of the
Cue 252 — 00:14:42,514 --> 00:14:45,354
problem. We do have to do that, but we also have to start thinking about what new
Cue 253 — 00:14:45,394 --> 00:14:48,774
libraries look like, what new frameworks look like, and how to migrate to them.
Cue 254 — 00:14:49,314 --> 00:14:53,034
Because these same technologies that are used for all of this can also be used to do
Cue 255 — 00:14:53,074 --> 00:14:56,754
the upgrades, right. So as a piece of advice for everyone, when we start to look at
Cue 256 — 00:14:56,794 --> 00:15:00,954
the micro view, the best thing anybody can do right now is just start upgrading all
Cue 257 — 00:15:00,994 --> 00:15:02,614
of the software that you have.
Cue 258 — 00:15:02,674 --> 00:15:04,114
It's going to get you closer to the edge.
Cue 259 — 00:15:04,154 --> 00:15:07,114
It's going to get you past all of the silent vulnerabilities that are that are
Cue 260 — 00:15:07,154 --> 00:15:09,574
getting found every day and not really reported.
Cue 261 — 00:15:10,154 --> 00:15:11,514
And at the same time, it's getting you new features.
Cue 262 — 00:15:11,554 --> 00:15:12,634
It's getting better performance.
Cue 263 — 00:15:12,674 --> 00:15:14,314
All of these libraries are moving fast.
Cue 264 — 00:15:14,354 --> 00:15:18,634
So if you're staying on older ones and hoping for LTS releases to back port
Cue 265 — 00:15:18,674 --> 00:15:22,554
vulnerability fixes, you're giving up years of development, years of feature work in
Cue 266 — 00:15:22,594 --> 00:15:25,314
those projects to and you're going to have to upgrade them someday.
Cue 267 — 00:15:25,354 --> 00:15:28,174
So why not today? It's kind of the moral of that story.
Cue 268 — 00:15:31,314 --> 00:15:33,034
Again mentioned before, it's not all scary, right?
Cue 269 — 00:15:33,074 --> 00:15:36,514
This stuff can be fun and it's going to be exciting once we figure out how to do it.
Cue 270 — 00:15:36,554 --> 00:15:38,994
And we get everyone starting to move at machine speed.
Cue 271 — 00:15:39,034 --> 00:15:42,394
Once everyone's doing that together, I think software is going to be a lot better.
Cue 272 — 00:15:42,434 --> 00:15:45,514
The industry is going to be a lot more fun, and we're going to be doing a lot more.
Cue 273 — 00:15:45,554 --> 00:15:47,694
We're all going to be shipping a lot more.
Cue 274 — 00:15:50,154 --> 00:15:53,554
I hope this quote right. Linus Torvalds, maintainer of the Linux kernel,
Cue 275 — 00:15:53,594 --> 00:15:59,814
on his battlecry to kernel maintainers everywhere, is that we never break userspace,
Cue 276 — 00:15:59,954 --> 00:16:01,914
and that makes sense down at that level of the stack.
Cue 277 — 00:16:01,954 --> 00:16:04,994
But just imagine you and your users, if you could ship software without having to
Cue 278 — 00:16:05,034 --> 00:16:07,794
worry about a decade of backwards compatibility issues.
Cue 279 — 00:16:07,834 --> 00:16:10,114
I think that's what we're going to get to machine readable change logs,
Cue 280 — 00:16:10,154 --> 00:16:13,354
where agents can take every one of those breaking changes without people having to
Cue 281 — 00:16:13,394 --> 00:16:14,534
worry about it.
Cue 282 — 00:16:17,844 --> 00:16:19,594
And we've done a bunch of experiments here.
Cue 283 — 00:16:19,634 --> 00:16:22,254
Agents are great at doing these upgrades.
Cue 284 — 00:16:22,474 --> 00:16:25,554
If you're on a nine year old version of some framework and ask the agent to do it
Cue 285 — 00:16:25,594 --> 00:16:27,994
all in one shot, it's probably going to make a mistake.
Cue 286 — 00:16:28,034 --> 00:16:31,274
But you wouldn't do that. You would ask the agent to do every single dot upgrade by
Cue 287 — 00:16:31,314 --> 00:16:34,594
itself over weeks, and then come back at the end.
Cue 288 — 00:16:34,634 --> 00:16:37,374
That's something an agent can do that a person would never do.
Cue 289 — 00:16:37,754 --> 00:16:40,594
People would just put this off until they absolutely have to,
Cue 290 — 00:16:40,634 --> 00:16:43,714
because the project goes end of life, and then probably just throw the thing out and
Cue 291 — 00:16:43,754 --> 00:16:46,254
rewrite it instead of actually ever doing this upgrade.
Cue 292 — 00:16:46,354 --> 00:16:50,794
But if you start thinking about this in smaller pieces, agents do an incredible job
Cue 293 — 00:16:50,834 --> 00:16:51,734
at this.
Cue 294 — 00:16:53,594 --> 00:16:55,554
We also prototyped a bunch of this ourselves.
Cue 295 — 00:16:55,594 --> 00:17:00,034
We've got a Java template library for generating HTML that we built in a couple of
Cue 296 — 00:17:00,074 --> 00:17:04,714
weekends, just with tokens. This gets rid of every class of vulnerability in every
Cue 297 — 00:17:04,754 --> 00:17:07,054
Java template library that we could find.
Cue 298 — 00:17:07,114 --> 00:17:08,794
It's completely backwards incompatible though.
Cue 299 — 00:17:08,834 --> 00:17:11,694
There's no way to click a single button to migrate to this.
Cue 300 — 00:17:12,074 --> 00:17:13,774
That was a benchmark that we set up.
Cue 301 — 00:17:13,954 --> 00:17:18,954
How fast can a Claude Sonnet 5 class agent with a single prompt of migrate me to
Cue 302 — 00:17:18,994 --> 00:17:21,414
this? Do it without breaking anything?
Cue 303 — 00:17:22,154 --> 00:17:25,474
It's not hard. This is kind of the pill that everyone's stuck on.
Cue 304 — 00:17:25,514 --> 00:17:28,294
These old frameworks is going to have to swallow.
Cue 305 — 00:17:28,394 --> 00:17:31,874
You can't stay on Spring Boot 2 for the next ten years, and hope that you're going
Cue 306 — 00:17:31,914 --> 00:17:34,174
to get enough back ports to keep yourself secure.
Cue 307 — 00:17:34,474 --> 00:17:36,634
So you're just going to have to figure out how to get agents to do all this
Cue 308 — 00:17:36,674 --> 00:17:38,174
migration work for you.
Cue 309 — 00:17:40,434 --> 00:17:44,014
Here are some of the stats. Most apps could be migrated for $9.
Cue 310 — 00:17:44,194 --> 00:17:49,314
The most expensive Java app that we could find on GitHub got migrated to this
Cue 311 — 00:17:49,354 --> 00:17:53,474
throwaway, completely incompatible library and fixed hundreds of vulnerabilities for
Cue 312 — 00:17:53,514 --> 00:17:57,914
$35, which is a lot of Sonnet 5 tokens, but it was still a lot cheaper than somebody
Cue 313 — 00:17:57,954 --> 00:18:01,014
doing this themselves over a couple of days.
Cue 314 — 00:18:04,114 --> 00:18:07,874
So, the takeaway to leave it on: This is going to be a lot of change coming at
Cue 315 — 00:18:07,914 --> 00:18:11,394
everyone very quickly. The only way out is actually it's going to be to upgrade
Cue 316 — 00:18:11,434 --> 00:18:14,794
faster, get used to upgrading, and then start to think about better software
Cue 317 — 00:18:14,834 --> 00:18:17,514
components that you can use without having to worry about these backwards
Cue 318 — 00:18:17,554 --> 00:18:19,414
compatibility guarantees anymore.
Cue 319 — 00:18:19,914 --> 00:18:23,134
Three months ago, this would have sounded crazy today.
Cue 320 — 00:18:23,354 --> 00:18:26,614
It's definitely doable. Three months from now, it's going to be even better.
Cue 321 — 00:18:28,514 --> 00:18:33,694
So yeah, this is probably the most exciting time to be in software in my career
Cue 322 — 00:18:33,954 --> 00:18:35,514
because we can start to think bigger.
Cue 323 — 00:18:35,554 --> 00:18:37,174
In addition to patching the planet.
Cue 324 — 00:18:37,234 --> 00:18:41,114
Figure out whole new frameworks, new operating systems, new libraries that aren't
Cue 325 — 00:18:41,154 --> 00:18:44,334
just more secure, they're faster and better and more fun at the same time.
Cue 326 — 00:18:45,314 --> 00:18:49,794
So I'll leave it here. The takeaway for everyone: start upgrading everything now.
Cue 327 — 00:18:49,834 --> 00:18:51,034
You're going to have to do it.
Cue 328 — 00:18:51,074 --> 00:18:55,034
Don't wait until there's a vulnerability on Hacker News telling you that you have to
Cue 329 — 00:18:55,074 --> 00:18:59,354
today. Plan it around your own schedule instead, and start thinking about what it
Cue 330 — 00:18:59,394 --> 00:19:00,954
looks like to rebuild the planet.
Cue 331 — 00:19:00,994 --> 00:19:02,514
What programming languages would be around?
Cue 332 — 00:19:02,554 --> 00:19:04,114
What libraries? What frameworks?
Cue 333 — 00:19:04,154 --> 00:19:07,114
What do you want to exist that doesn't exist yet?
Cue 334 — 00:19:07,154 --> 00:19:08,634
Because you can start to do that now.
Cue 335 — 00:19:08,674 --> 00:19:10,654
We can all think a lot bigger and bolder.
Cue 336 — 00:19:10,954 --> 00:19:13,394
And if you get all this right, you can put all of your tokens into shipping your own
Cue 337 — 00:19:13,434 --> 00:19:16,934
product instead of patching vulnerabilities and upgrading everything every day.
Cue 338 — 00:19:18,594 --> 00:19:20,434
I gave a bunch of time for questions.
Cue 339 — 00:19:20,474 --> 00:19:24,774
Now we have a microphone. What's that?
Cue 340 — 00:19:26,714 --> 00:19:27,974
How much time do we have?
Cue 341 — 00:19:35,774 --> 00:19:37,554
Perfect. Yeah.
Cue 342 — 00:19:39,694 --> 00:19:41,134
It's small enough. You can probably just talk.
Cue 343 — 00:19:41,174 --> 00:19:41,974
I don't know.
Cue 344 — 00:19:43,254 --> 00:19:48,094
Where are you at? On recommending, like, a cooldown on an update to a library.
Cue 345 — 00:19:48,134 --> 00:19:50,614
Obviously, if everyone does it, it's no good for anybody.
Cue 346 — 00:19:50,654 --> 00:19:51,894
What's the practical takeaway?
Cue 347 — 00:19:51,934 --> 00:19:57,734
And then a follow up. How often are you recommending that people rebuild just the
Cue 348 — 00:19:57,774 --> 00:20:01,014
tiny part of the library that they're doing in-house, instead of like,
Cue 349 — 00:20:01,054 --> 00:20:04,894
exposing themselves surface area wise, to a larger charting library?
Cue 350 — 00:20:04,934 --> 00:20:06,974
You know, remote code execution.
Cue 351 — 00:20:07,014 --> 00:20:10,794
Yeah. So two questions there. The first was cooldown periods,
Cue 352 — 00:20:11,094 --> 00:20:15,334
which is a common tactic for kind of waiting out the malware attacks that are also
Cue 353 — 00:20:15,374 --> 00:20:18,854
happening at the same time. Security teams are in a tough spot right now where you
Cue 354 — 00:20:18,894 --> 00:20:21,974
have to be ready to upgrade for zero days every single day, but also if you upgrade,
Cue 355 — 00:20:22,014 --> 00:20:24,274
there might be malware in the thing that you're grabbing.
Cue 356 — 00:20:24,494 --> 00:20:26,774
Cooldown periods can work. It kind of depends how long you want to wait,
Cue 357 — 00:20:26,814 --> 00:20:30,774
but at the same time, you could also be sitting on a known that you have to grab.
Cue 358 — 00:20:30,814 --> 00:20:35,934
So it's it can work fine, but you then end up in the position of reviewing all of
Cue 359 — 00:20:35,974 --> 00:20:38,174
that stuff to to allow things through.
Cue 360 — 00:20:38,214 --> 00:20:40,874
If you do need to get one of these critical patches through.
Cue 361 — 00:20:41,214 --> 00:20:42,574
So it's important a lot of people are doing it.
Cue 362 — 00:20:42,614 --> 00:20:46,714
GitHub has built in features PyPI and pip just added one.
Cue 363 — 00:20:47,094 --> 00:20:49,954
I've heard seven days. Some people are doing 14 days.
Cue 364 — 00:20:50,134 --> 00:20:51,214
That's the trade off you have to make.
Cue 365 — 00:20:51,254 --> 00:20:55,034
How long do you want to wait versus how long or how fast you actually have to patch?
Cue 366 — 00:20:55,454 --> 00:20:58,774
It starts to get really hard when you're cherry picking of don't take anything
Cue 367 — 00:20:58,814 --> 00:21:02,474
except for this one, but that one has transitive dependencies that are also newer,
Cue 368 — 00:21:02,974 --> 00:21:06,194
so it's more complicated than it sounds at first.
Cue 369 — 00:21:06,614 --> 00:21:11,014
The second one was what about just forking and copying and parts of a library that
Cue 370 — 00:21:11,054 --> 00:21:13,074
you need rather than the entire thing?
Cue 371 — 00:21:13,974 --> 00:21:17,214
It's hit or miss. You know, there are some kitchen sink frameworks where you want
Cue 372 — 00:21:17,254 --> 00:21:21,474
one function and you pull in a million of them at the same time.
Cue 373 — 00:21:22,334 --> 00:21:26,454
The other side of it is the JavaScript ecosystem, where their entire libraries that
Cue 374 — 00:21:26,494 --> 00:21:31,174
are just one function, you can copy that stuff yourself, but then you're on the hook
Cue 375 — 00:21:31,214 --> 00:21:34,814
for maintaining it. You're going to lose all the metadata about knowing where the
Cue 376 — 00:21:34,854 --> 00:21:38,854
vulnerabilities are. If there is one in that library and you forked it and renamed
Cue 377 — 00:21:38,894 --> 00:21:42,014
it, and you only have a part of it, you're losing all of the machine readable
Cue 378 — 00:21:42,054 --> 00:21:44,974
metadata that would tell you whether that's in the part that you forked and forgot
Cue 379 — 00:21:45,014 --> 00:21:49,054
about. I know Mitchell Hashimoto is a big fan of that approach and stuff and keeps
Cue 380 — 00:21:49,094 --> 00:21:51,454
promoting it. I wouldn't recommend it.
Cue 381 — 00:21:51,494 --> 00:21:52,934
It's too scary. It's too hard.
Cue 382 — 00:21:52,974 --> 00:21:55,554
You're going to end up maintaining more code than you realize.
Cue 383 — 00:21:58,014 --> 00:22:01,534
Yeah, most languages already have decent enough to where you're not pulling every
Cue 384 — 00:22:01,574 --> 00:22:02,694
single thing into your library.
Cue 385 — 00:22:02,734 --> 00:22:05,714
It's only pulling in the things that you actually need at the end of the day.
Cue 386 — 00:22:11,494 --> 00:22:14,774
Hey, yeah, I want to kind of build on that because there was a talk earlier today
Cue 387 — 00:22:14,814 --> 00:22:17,094
from Geoff Huntley. I was looking around to see if he's here.
Cue 388 — 00:22:17,134 --> 00:22:20,614
That was basically like stop using anyone else's dependencies.
Cue 389 — 00:22:20,654 --> 00:22:24,294
And it's I mean, I get that we don't know the answer right now,
Cue 390 — 00:22:24,334 --> 00:22:26,894
right? Both of those things. Both.
Cue 391 — 00:22:26,934 --> 00:22:30,814
But I think the, the metadata piece of like, now, I don't even know that I copied
Cue 392 — 00:22:30,854 --> 00:22:32,894
something or like, I got the agent to build something.
Cue 393 — 00:22:32,934 --> 00:22:33,734
So,
Cue 394 — 00:22:35,134 --> 00:22:37,734
you know, a lot of work. I guess I was just a comment on that one.
Cue 395 — 00:22:37,774 --> 00:22:42,894
A lot of work has gone in recently to, you know, are we actually using this part of
Cue 396 — 00:22:42,934 --> 00:22:44,894
the library? Do we need to make an update?
Cue 397 — 00:22:44,934 --> 00:22:49,054
You know, and it sounds like you're advocating just keep everything up to date
Cue 398 — 00:22:49,094 --> 00:22:51,374
because that's actually the easiest way to manage it.
Cue 399 — 00:22:51,414 --> 00:22:53,674
Yeah. Is there.
Cue 400 — 00:22:55,494 --> 00:22:58,394
This is going to be the most ridiculous question, but now that I got to it,
Cue 401 — 00:22:59,054 --> 00:23:02,894
how do you how do you manage that with like, auditors GRC, like the kinds of people
Cue 402 — 00:23:02,934 --> 00:23:06,894
that are that are driving these decisions tend to be ten years plus outdated in the
Cue 403 — 00:23:06,934 --> 00:23:08,094
way that they think about them.
Cue 404 — 00:23:08,134 --> 00:23:08,934
So,
Cue 405 — 00:23:10,094 --> 00:23:14,214
what kind of advice do you have for kind of managing, not the actual security,
Cue 406 — 00:23:14,254 --> 00:23:17,814
which I totally get, but the process around that and the process of working with
Cue 407 — 00:23:17,854 --> 00:23:19,214
your customers and that sort of thing.
Cue 408 — 00:23:19,254 --> 00:23:23,714
Yeah. Yeah. I'd love to go have some debate over beers with Geoff on that one.
Cue 409 — 00:23:23,774 --> 00:23:27,774
You know, I think I don't think everyone should be rewriting all of their
Cue 410 — 00:23:27,814 --> 00:23:29,534
dependencies, right? I know you can.
Cue 411 — 00:23:29,574 --> 00:23:33,414
Sometimes it's fun, but I don't think we're ever going to get to a world.
Cue 412 — 00:23:33,454 --> 00:23:36,614
Or at least that's a terrible thing to say in today's world.
Cue 413 — 00:23:36,654 --> 00:23:39,454
I don't think in the next 18 months we're going to get to world where even AI is
Cue 414 — 00:23:39,494 --> 00:23:42,334
writing perfect software, and I'll back that up a little bit.
Cue 415 — 00:23:42,374 --> 00:23:47,334
Like Opus 5.5 today is writing better software than most engineers,
Cue 416 — 00:23:47,374 --> 00:23:52,634
almost every engineer. But the gaps and the bugs come in between the sessions.
Cue 417 — 00:23:53,094 --> 00:23:55,774
When you look at model performance, it's getting better and better every day,
Cue 418 — 00:23:55,814 --> 00:23:58,734
but the context windows aren't getting any longer, right?
Cue 419 — 00:23:58,774 --> 00:24:03,174
We've been kind of the 400K, 1 million token contexts we know for the last couple of
Cue 420 — 00:24:03,214 --> 00:24:07,294
years. I'm not a hardware expert, but I think that's a harder needle to move than
Cue 421 — 00:24:07,334 --> 00:24:12,654
the model performance itself. And until you can get to a spot where you can load
Cue 422 — 00:24:12,694 --> 00:24:16,734
your entire program into the context window, I don't see how you avoid those seams
Cue 423 — 00:24:16,774 --> 00:24:19,194
and the gaps. And I ran it today.
Cue 424 — 00:24:19,614 --> 00:24:21,774
It was perfect code. I added a feature tomorrow.
Cue 425 — 00:24:21,814 --> 00:24:23,614
It didn't have the entire context in there.
Cue 426 — 00:24:23,654 --> 00:24:27,494
And now the vulnerability or the bugs show up in those gaps in those seams.
Cue 427 — 00:24:27,534 --> 00:24:31,134
So I think we're still in a world of, you know, old law.
Cue 428 — 00:24:31,174 --> 00:24:33,114
Many eyes make all bugs shallow.
Cue 429 — 00:24:33,334 --> 00:24:35,774
I think we'll all be better if everyone is pointing their agents at one thing,
Cue 430 — 00:24:35,814 --> 00:24:39,274
rather than everybody ignoring them and building their own versions.
Cue 431 — 00:24:39,334 --> 00:24:41,374
I do think open source will probably change a little bit though,
Cue 432 — 00:24:41,414 --> 00:24:44,034
right? Like, you're not going to vibe code a database.
Cue 433 — 00:24:44,974 --> 00:24:49,174
You're not going to build your own nginx or proxy or any of these security critical
Cue 434 — 00:24:49,214 --> 00:24:53,294
libraries. That stuff becomes even more important, and you can build better ones and
Cue 435 — 00:24:53,334 --> 00:24:56,454
make them even better and faster if everybody's collaborating on those things
Cue 436 — 00:24:56,494 --> 00:25:01,414
together. The stuff in the middle, you know, the ORMs and all of these frameworks
Cue 437 — 00:25:01,454 --> 00:25:04,734
that you pull in because everyone else is, but you end up fighting against them half
Cue 438 — 00:25:04,774 --> 00:25:07,614
the time anyway. You know, agents are great at writing SQL.
Cue 439 — 00:25:07,654 --> 00:25:10,054
You don't need an ORM anymore to help you write SQL.
Cue 440 — 00:25:10,094 --> 00:25:12,974
I think a lot of those things in the middle are going to become less important,
Cue 441 — 00:25:13,014 --> 00:25:15,754
and you're probably better off just doing some of that stuff raw,
Cue 442 — 00:25:16,094 --> 00:25:19,674
rather than building your own ORM or something like that.
Cue 443 — 00:25:21,454 --> 00:25:22,814
Your second question, I forgot it.
Cue 444 — 00:25:22,854 --> 00:25:24,954
No, sorry. I went yeah. Oh, yeah.
Cue 445 — 00:25:25,294 --> 00:25:28,774
You know, it's for the most part, they love seeing no vulnerabilities.
Cue 446 — 00:25:28,814 --> 00:25:32,194
The CVE counts, for a while. They're kind of garbage and made up anyway,
Cue 447 — 00:25:32,254 --> 00:25:34,974
but nothing makes them happier than seeing a zero there.
Cue 448 — 00:25:35,014 --> 00:25:37,854
So for the most part, there's not really a lot of friction.
Cue 449 — 00:25:37,894 --> 00:25:41,094
They kind of like it. It depends what industry you're in the like,
Cue 450 — 00:25:41,134 --> 00:25:44,774
if you have change review boards and stuff like this, or a bunch of execs in a room
Cue 451 — 00:25:44,814 --> 00:25:48,334
have to decide whether it's okay to upgrade that thing, you probably have to go
Cue 452 — 00:25:48,374 --> 00:25:51,374
solve that culturally. And that's that's what you're going to have to do to start
Cue 453 — 00:25:51,414 --> 00:25:53,074
moving at machine speed as well.
Cue 454 — 00:25:54,134 --> 00:25:54,934
Yeah.
Cue 455 — 00:25:57,774 --> 00:25:58,834
In the back.
Cue 456 — 00:26:04,014 --> 00:26:08,774
Sad that you need to do this talk and tell people that, you know,
Cue 457 — 00:26:08,814 --> 00:26:16,014
updating their software is a thing, you know, and, it's even ironic
Cue 458 — 00:26:16,054 --> 00:26:20,014
when you see governments still running PHP, especially the old PHP.
Cue 459 — 00:26:20,054 --> 00:26:22,234
Like, why? Why are you there?
Cue 460 — 00:26:23,574 --> 00:26:29,094
Do you think maybe it's like, a cultural thing and we finally need to start doing
Cue 461 — 00:26:29,134 --> 00:26:34,334
what we were supposed to be doing for the last 15 years, since AI allows us,
Cue 462 — 00:26:34,374 --> 00:26:37,614
like. Yeah, I think it's it's both sides there.
Cue 463 — 00:26:37,654 --> 00:26:40,654
Right? You know, the question is, why do we have to give this talk about upgrading?
Cue 464 — 00:26:40,694 --> 00:26:42,894
It's the most boring thing in the world is depressing.
Cue 465 — 00:26:42,934 --> 00:26:47,574
But yeah, the amount of customers I had to talk to that were begging for back ported
Cue 466 — 00:26:47,614 --> 00:26:50,634
PHP vulnerability fixes to stuff like PHP five.
Cue 467 — 00:26:51,024 --> 00:26:56,014
You can do it. But if you're still running PHP five, there's at least 30 other even
Cue 468 — 00:26:56,054 --> 00:26:58,854
worse vulnerabilities that have been known for the last decade that you somehow
Cue 469 — 00:26:58,894 --> 00:27:02,414
didn't bother worrying about. You're just worried about this one because it was in
Cue 470 — 00:27:02,454 --> 00:27:06,594
the news, you know, which is a terrible way to do security and risk management.
Cue 471 — 00:27:06,694 --> 00:27:07,854
But I think it's the same thing you talked about.
Cue 472 — 00:27:07,894 --> 00:27:09,594
Upgrading is hard. It's scary.
Cue 473 — 00:27:09,974 --> 00:27:11,894
If you upgrade stuff, it might break.
Cue 474 — 00:27:11,934 --> 00:27:14,894
But you know what's great at doing these upgrades and looking at every single line
Cue 475 — 00:27:14,934 --> 00:27:16,694
and making sure you didn't miss any of these steps.
Cue 476 — 00:27:16,734 --> 00:27:21,094
It's AI, but the same organizations that haven't upgraded are also probably the
Cue 477 — 00:27:21,134 --> 00:27:23,974
slowest to start adopting these tools, which is where the friction starts to come
Cue 478 — 00:27:24,014 --> 00:27:28,974
in. Yeah, but I think the cause of and solution to all of our problems,
Cue 479 — 00:27:29,014 --> 00:27:32,994
I guess, is kind of to sum it up all the way in the back.
Cue 480 — 00:27:39,954 --> 00:27:42,874
Thoughts on the EU Cyber Resilience Act.
Cue 481 — 00:27:42,914 --> 00:27:49,874
Yeah, it's very long and I think a lot of it is still to be
Cue 482 — 00:27:49,914 --> 00:27:51,274
determined how it's going to play out.
Cue 483 — 00:27:51,314 --> 00:27:55,434
Right. You know, it's not as bad as it looked like it was going to be from an open
Cue 484 — 00:27:55,474 --> 00:28:00,074
source perspective, some of the first couple of drafts where there was just a
Cue 485 — 00:28:00,114 --> 00:28:03,794
fundamental misunderstanding of how open source works and how open source licenses
Cue 486 — 00:28:03,834 --> 00:28:08,714
work. Some of the first couple of drafts actually put legal liability on open source
Cue 487 — 00:28:08,754 --> 00:28:11,954
maintainers, whether or not they have any commercial relationship with the people
Cue 488 — 00:28:11,994 --> 00:28:15,954
using it. If you put something up on GitHub, some bank in Europe uses it and gets
Cue 489 — 00:28:15,994 --> 00:28:19,394
hacked, it's your fault. Somehow that's been softened.
Cue 490 — 00:28:19,434 --> 00:28:21,274
It hasn't been completely removed.
Cue 491 — 00:28:21,314 --> 00:28:28,074
I think it's going to take a couple examples and actual probably judges to decide
Cue 492 — 00:28:28,114 --> 00:28:30,714
how to actually apply it in some of these cases, but for the most part,
Cue 493 — 00:28:30,754 --> 00:28:33,434
it's getting people to pay attention to what's in their software,
Cue 494 — 00:28:33,474 --> 00:28:34,614
which is a good thing.
Cue 495 — 00:28:38,934 --> 00:28:40,194
Thank you everyone.
.tessl-plugin
skills
talk-arcolano-autonomy-and-trust
talk-aroussi-decision-context
talk-bevin-agent-assisted-instrument-design
talk-boeckeler-context-lifecycles
talk-braun-context-graphs
talk-carmont-evolving-agent-harnesses
talk-chanezon-gordon-agent-experience-evaluation
talk-cox-secure-agentic-workflows
talk-cunniffe-reusable-agent-platform
talk-davis-agent-tool-policies
talk-debois-concurrent-delivery
talk-dzialets-team-automation
talk-fahmy-harness-api-design
talk-finster-continuous-delivery
talk-fowler-regenerative-software
talk-gavrilescu-factory-architecture
talk-graciolli-amin-verification-aware-programming
talk-gruenbaum-reverse-graph-engineering
talk-hassan-agentic-engineering
talk-hossain-enterprise-agent-adoption
talk-huntley-patch-ready-factories
talk-jain-verifying-intent
talk-katsioloudes-montei-agentic-lifecycle
talk-knox-harness-engineering
talk-kovac-deterministic-memory
talk-lee-perkins-code-readiness
talk-lorenc-machine-speed-security
talk-lubken-software-of-one
talk-meijer-software-specialization
talk-meppiel-versioned-agent-contracts
talk-moss-concurrent-agent-risk
talk-myers-understanding-agent-generated-code
talk-ocytko-scaling-agentic-engineering
talk-oza-automating-design-quality
talk-panzica-agentic-sdlc
talk-pari-patel-agent-awareness
talk-podjarny-trustworthy-software-factories
talk-poulter-factory-improvement
talk-pratik-patel-refined-rag-data
talk-re-ferre-development-platforms
talk-sadogursky-factory-of-factories
talk-shah-training-agent-skills
talk-sharma-mohapatra-evolving-intent
talk-singhal-operating-agents-at-scale
talk-stack-incident-dossiers
talk-tyler-team-coding-benchmarks
talk-wotherspoon-agent-execution-environments
talk-yegge-ai-powered-factories
talk-zuber-leadership-through-coding