# Democratizing Security for the Agentic Web with Cloudflare CTO Dane Knecht

No Priors: AI, Machine Learning, Tech, & Startups · 2026-10-23

<https://nopriors.podhood.com/d90f9911-a1fc-41a2-8a50-58bb566bcd26>

Cloudflare CTO Dane Knecht tells Sarah Guo that AI has made vulnerability discovery relentless—once-yearly exploits now surface weekly, often from high schoolers with agents—pushing over 50% of engineering time to security for the next 18 months. He argues patching 70 years of software demands a redesigned development lifecycle and smaller attack surface, and that zero-trust architecture leaves enterprises better positioned. Calling the web a 'messy middle' where humans and agents share rails, he details Cloudflare's monetization gateway with agent wallets and a stablecoin plus Signals with OpenAI for CPM-style creator payments. He covers Cloudflare OS, a restructuring that folded a three-team email workflow into one tool, token budgeting, 1,000 interns, and four to five red teams.

## Questions this episode answers

### Who is behind the surge in AI-driven vulnerability discovery?

Dane Knecht says many recent large vulnerabilities weren't from nation-states but from high school kids — Sarah Guo dubs them "agent kiddies" — who point agents at targets like capture-the-flag contests, earning sizable bug bounties. This gives the masses tools nation-states previously had.

[5:26](https://nopriors.podhood.com/d90f9911-a1fc-41a2-8a50-58bb566bcd26?t=326000)

### What is Cloudflare's philosophy on token budgeting for employees?

Dane Knecht says asking humans to pick models per task is silly and routing should become intelligent. Cloudflare has no token leaderboard and grades people on output, but sends notices at $300 a day, escalates to a manager at $500 and a VP at $1,000.

[27:25](https://nopriors.podhood.com/d90f9911-a1fc-41a2-8a50-58bb566bcd26?t=1645000)

### Why did Cloudflare hire a thousand interns?

Dane Knecht explains that two years ago Cloudflare had about 200 interns who were the first fully AI-native workers, and managers with those interns adopted AI faster. So the company hired 1,000 interns for its roughly 5,000-person company to bring in AI-native ideas and habits.

[29:55](https://nopriors.podhood.com/d90f9911-a1fc-41a2-8a50-58bb566bcd26?t=1795000)

### Should cybersecurity AI models be open or closed?

Dane Knecht argues that over time cyber models all have to be open, because bad actors will get access anyway and top companies can too — keeping them closed just excludes the mass population of defenders who most need the capabilities.

[38:49](https://nopriors.podhood.com/d90f9911-a1fc-41a2-8a50-58bb566bcd26?t=2329000)

## Key moments

- **[0:00] Intro**
  - [0:00] Cloudflare CTO Dane Knecht: "AI has become the tool that's causing this security apocalypse — and also the tool that can help us solve it"
- **[0:55] Top priorities**
  - [2:49] Dane Knecht predicts over 50% of engineering time will go to security within 18 months as CVEs now surface weekly, not yearly
  - [3:24] A key protecting all TLS and WebPKI on the internet was secured by the number four — and OpenAI just found the NIST curve seeds from 2000
  - [5:00] The biggest new vulnerabilities aren't nation-states — they're high school kids pointing AI agents at targets, says Cloudflare's CTO
- **[5:03] Attackers & spend**
  - [6:12] Air-gap-everything doom vs zero-trust optimism: Dane Knecht argues industry rivals will unite to protect critical infrastructure
  - [7:53] Dane Knecht says emergency will push security token spend to match engineering spend within 18 months, then productivity rebounds
- **[8:39] Secure software**
  - [9:14] Cloudflare is patching 70 years of software, and the agent development lifecycle must bake in security at every step
  - [12:05] Cloudflare's internal 'Codex' distills RFCs into ten commandments that AI agents follow during every code review
- **[12:54] The messy middle**
  - [13:19] The web's 'messy middle': humans and agents ride the same rails, and Dane Knecht says we'll end up with parallel human and machine webs
  - [15:01] Dane Knecht predicts agents will pay a site's CPM rate to read content, finally forcing real micropayment rails onto the internet
- **[17:27] Monetization**
  - [17:38] Cloudflare launched a content monetization gateway with agent wallets and a stablecoin, plus the Signals project with OpenAI
- **[20:07] Cloudflare OS**
  - [20:18] Cloudflare OS is the company's internal AI platform that lets any employee build tools without having to think about security
- **[24:07] Restructuring**
  - [24:19] Cloudflare's AI-first restructuring replaced a three-team email process with one tool where a PM describes a cohort and the manager clicks approve
- **[27:11] Token budgeting**
  - [27:25] Cloudflare's token budgeting: no leaderboard, but a $300-a-day spend triggers an email, $500 goes to your manager, $1,000 to a VP
- **[29:32] Intern strategy**
  - [29:41] Cloudflare hired 1,000 interns for its 5,000-person company after seeing AI-native interns made their managers adopt AI faster
- **[33:33] Hiring & curiosity**
  - [33:45] Dane Knecht: if a hire isn't using AI at home in their personal life, "that's a pretty big red flag"
- **[35:23] Red teams**
  - [35:42] Cloudflare is going from one red team to four or five, betting engineers will automate patching once vulnerabilities get too big
  - [39:09] Open vs closed AI models in cybersecurity: Dane Knecht says defensive capabilities must eventually all be open because attackers get access anyway
  - [39:58] Cloudflare gives every free customer protection against major vulnerabilities like the React bug, calling security a public utility

## Speakers

- **Sarah Guo** (host)
- **Dane Knecht** (guest)

## Topics

AI Security & Infrastructure

## Mentioned

Cloudflare (company), OpenAI (company), Stripe (company), Cloudflare OS (product), Muse (product), Signals (product)

## Transcript

### Intro

**Dane Knecht** [0:00]
AI has become the tool that's causing this kind of security apocalypse. It is also the tool that can help us solve it faster as well. It went from "how do you get everyone just using AI" to now all of a sudden everyone's racing to figure out what AI security means.

**Sarah Guo** [0:11]
I've heard serious technical people say things like, "Ear gap everything important," and we're like, not going to be able to trust the internet for a while. And we just assume all the critical infrastructure and government services fall down at some point.

Is that your expectation for what happens?

Dane, I'm so excited to do this with you.

**Dane Knecht** [0:37]
Yeah, thank you. Thanks for coming here too.

**Sarah Guo** [0:39]
I'm sure I'm not the first person to make this joke, but your last name, Knecht, that's pretty good for the Cloudflare value prop.

**Dane Knecht** [0:47]
Yeah, yeah. There's been some jokes about renaming the show next year, so.

**Sarah Guo** [0:52]
So with everything that Cloudflare is doing, what are your top few priorities as CTOright now?

### Top priorities

**Dane Knecht** [0:59]
Well, I mean, this past year, it feels like 10 years compacted into one. So those priorities are constantly shifting. But, you know, it all goes back to our mission of helping build a better internet. And that really hasn't changed over the last 15 years, where we want to make sure that the internet's secure, that everyone has access to be able to get their content online easily and scale it up.

And just, you know, over the arc of the AI changes in the past year, you know, it went from, you know, how do you get everyone just using AI, how do you make it easy for them to deploy and their agents to deploy, to like most technical transitions, security is kind of an afterthought.

And now all of a sudden everyone's racing to figure out what AI security means. And it means things for, you know, different sides view it kind of differently. You know, one side is, you know, as a company, is how do you keep your data safe from AI leaking the data out?

But it's also how do you keep people from getting in as well too. And, you know, over the last, I would say, three to six months, just the number of vulnerabilities that are getting reported in CVEs has just gone up exponentially.

You know, previously on the internet, we would see these log for J moments, or these shell shock, or type of incidents maybe once a year. And it would be all hands on deck internally at Cloudflare to make sure how we protect not just, you know, our enterprise customers, but also all of our free customers, because we, you know, view it as kind of a public good of making sure that the internet stays safe.

To these things are now happening weekly. Not all of them are publicly reported on, but there are things that, you know, are as serious as the, like, the log for J type things, but just happening at a greater frequency.

You know, and I expect that, you know, through the rest of this year and in the next year, by the next 18 months, the majority of teams' time is going to be spent on security. You know, I think there was a stat that was put up earlier today that during one of our keynotes where it was showing that, I think, over 50% of engineering time is going to be start to be spent on security.

And I don't think that will last forever. I think that will be a short transition. You know, at some point, the AI will have surveyed the majority of the code and found most of the bugs or found most of the shortcuts.

You know, there's one that's kind of coming out today and has been started to be blogged about and tweeted about, where one of the keys on the internet is that protects all TLS and WebPKI, which secures all your connections,

was secured by the number four. Literally, they basically used, you know, about cryptography, I mean, it's all about randomness,right? You know, that's why Cloudflare has our lava lamps is so that you can get good randomness. And when a key out there is they use the seed, they use the initial thing to kind of create it with a single number four, it's going to be reverse engineered.

You know, and AI can just do all these things, you know, it doesn't sleep. It's not like the one engineer you had that was red teaming things. You know, last week, I think OpenAI found the seeds for the NIST curves that were back in 2000 that were released.

So just, you know, these things are just moving so fast and the bar is going up and the models are just getting so good at, you know, cybersecurity. You know, MITSES was just kind of the tip of the iceberg there where, you know, and since then, it's not just been the models getting better, it's the harnesses.

It's how people know how to, how humans know how to use them and use those frontier models in order to find those exploits.

**Sarah Guo** [5:00]
Given your, you know, visibility, what do you know about the humans behind, you know, these attacks and harnessing these models and taking the volume of vulnerability discovery up so quickly? Like, are they, you know, nation-state actors? Are they script kiddies?

### Attackers & spend

**Sarah Guo** [5:19]
Like, what can you say about

how the attacker landscape has changed?

**Dane Knecht** [5:26]
Well, I actually think for some of these, some of these large vulnerabilities that have come out more recently, actually, haven't been nation-states. They've been, you know, high school kids even that are.

**Sarah Guo** [5:39]
We're going to call them agent kiddies now.

**Dane Knecht** [5:40]
Yeah, yeah. That are just spending their time, you know, like they used to do kind of capture the flag, security contest. They're pointing their agents at an area and doing that. You know, there's also, you know, and there's financial reward for them.

There's bug bounties that they get paid out, you know, sizable amounts. But it's given the tools that honestly the nation-states have had in some forms for much longer, either through agents or just through, you know, sure number of people that were working on the problems, giving it to the masses.

**Sarah Guo** [6:12]
This is the first time in 15 years I've heard serious technical people say things like, "Well, we should just like air gap everything important and we're like, you know, not going to be able to trust the internet for a while and we just assume all the critical infrastructure and government services fall down at some point."

Is that your expectation for what happens?

**Dane Knecht** [6:35]
Humans, we are pretty adaptable and we can move fast when we need to. And, you know, I've seen the industry kind of partner together. When things like this happen, people that, you know, usually think of competitors are no longer competitors.

We are just literally protecting critical infrastructure, as you said. And I think when those kind of critical moments happen, people will rise to that challenge. So I'm a little less doomsday. I do think that a lot of the things that we've done over the past, you know, five to ten years, moving away from a kind of a castle-and-moat infrastructure for security to moving to more of a zero trust does put a lot of companies in a better position where, you know, you're not saying this office is completely air gapped.

You're saying, you know, these devices, these flows are secured in a different way and everything's going to be tied in.

**Sarah Guo** [7:28]
What is it going to take for organizations to do what you're describing? Like 50% of engineers do not work on securityright now. I had Greg Brockman from OpenAI on the No Priors podcast last week, and he said that he thought in 2027, people are going to spend as much money on tokens for security as engineering.

Like you're saying, very related idea. What is going to drive that shift?

**Dane Knecht** [7:58]
Just emergency,right? So you're going to move forward all your security spend over the next 18 months, both from the red team side, which is more of the offensive, trying to actively try to exploit things, as well as the defensive side, how you prepare things.

But I don't think that's going to last forever. You know, I see a world in like 18 months where, you know, new software productivity goes way up because, you know, all those resources are back focused on that. But there's definitely this period where security needs to be on everybody's mind.

**Sarah Guo** [8:36]
What are some of the key elements of getting to a better place in terms of overall software and internet security? I think people discover, of course, like that there are libraries and services everyone uses that had vulnerabilities that took effort to discover and now do not.

### Secure software

**Sarah Guo** [8:55]
And I think we'll hit some of those. What else is important there? Because part of the concern that I have is the rate of software production is also going up a great deal,right? So to some degree, we got to go like patch everything that we have already made and then keep up and get ahead of the pace of delivery.

**Dane Knecht** [9:14]
Yeah. Also, I mean, for one, we'reright now, we're patching 70 years of software out there. So some of it is just catch up. And then on a go-forward basis, you know, I think this, you know, the software development lifecycle, the agent development lifecycle is going to need to change.

And, you know, luckily the same way that we can use securityright now for red teaming and things like that, it'll just be built into the process every step of the way.

And I think software will just get more secure over time.

**Sarah Guo** [9:47]
What have you learned about building secure production software internally already that you think people can learn from? Like what does that mean? Like shift left and like can we, and there are many theories of how to do this, including, you know, broader use of formal verification, using agents for security in the development process.

Like what do you think is going to create better software from first delivery?

**Dane Knecht** [10:16]
I kind of go back to first principles of theirs. You want to reduce your attack surface area. The easiest thing you can do is to give less places for there to be vulnerabilities. So, you know, looking at all your software and saying, how can I simplify this?

How can I reduce the footprint, the code? And then, you know, beyond that, I mean, things like formal verification, like I think things that, you know, safer programming languages and things that are kind of built into the runtimes that make everything safer will come.

But, you know, I think it just starts with, and, you know, AI, at least, you know, the early AI is very verbose. It likes to create complexity. So, you know, AI in a sense of what we've been dealing with has actually not been kind of helping build a more secure stuffright now, but I think it will get there.

And I think the, you know, that's where humans still come in and human judgment comes in and making sure that we design the software in a way that can stand on its own securely.

**Sarah Guo** [11:17]
That was actually going to be my reaction when you say we want to reduce the surface area. I'm like, "Oh, yes, of course, in principle." And that's not the natural inclination of today's models,right? And like, you know, I ask you to do something and I'm like, "Well, like with one-tenth the complexity, please try again."

Is that the role of engineering and security engineering today? Is that something that's going to be solved at the model level?

**Dane Knecht** [11:43]
The way that humans did it before agents, it needs to be solved kind of at all levels. You need to train the engineers that are writing the code, the product engineers. You need the security teams still need to be involved in making sure that they're getting, you know, checks and balances, second set of eyes on things.

You need to build in processes. We've gone to

defining things a lot more internally on how we build at Cloudflare, writing down our principles. We have a document called the Codex, which is kind of like,

you know, a buildup of a lot of RFCs that the teams have built for it. Here's how you build secure software at Cloudflare. Here's how you build available software at Cloudflare. Here's our best practices. And then it kind of gets bulled up into some 10 commandments that the agents know how to follow and use in every part of the review process that can help enforce those things.

So, you know, you have agents out there that are, you know, just writing a lot of the slop. And then you have deeper down in the factory, you have people that are, you know, doing the performance runs and the security runs.

And just like how the company operated before, I think you need those adversarial reviews.

**Sarah Guo** [12:50]
When you zoom out a little bit and just look at overall traffic on the internet and how much of it has become agent-driven over the last year, I think it's outpaced most people's expectations. What has updated in your worldview of like what the web a year from now looks like, either from an economic model perspective or a security or requisite infrastructure perspective?

### The messy middle

**Dane Knecht** [13:19]
Yeah. I mean, I kind of think of it as we're in the messy middleright now where humans and agents are riding the same rails on the internet, where, you know, agents are trying to parse a human web when that's just, you know, one is just inefficient, you know, and was designed for humans that could do things at different speeds, you know, going to a website and being able to read and, you know, see the ads and participate in the more of the full economic model versus an agent that now can go to a thousand web pages in the same period of time, which is great.

It's just you need to find a new economic model to support that.

**Sarah Guo** [13:58]
Do we get parallel webs? Is there like a human web and a machine web? Like everybody's trying to do, what's it called? AEO now?

**Dane Knecht** [14:07]
Yeah. I mean, yes, I think we do. I think we find ways to deliver the content and meet agents and humans where they are, as opposed to trying to meet them in the same place. Andright now there's like translation layers.

You know, there's a lot of companies that like, you know, help you, you know, an agent view a web page. These are just all extra, you know, burning of tokens for things that probably don't need to be, need to happen.

You know, the website usually knows what data. And, you know, we've had things like this, you know, we've had APIs, you know, ways that computers talk to each other for years. It's just, you know, APIs only covered a small fraction of the information on the internet.

And agents have a thirst to, you know, understand everything. And so they're going out and trying to explore the human internet. And, you know, we need to make it easier for them. We need to make it, you know, less costly for the site owners.

And we need to find a way for them to get compensated. You know, I think it's in a world where you can get very simple, you know, most like it's like a content website, like a Substack, you know, they know what their CPM is for how much money they should be making on ads for that site.

And so if an agent's reading that site, I think there is a world where you get to the point where an agent can decide whether or not they want to read the content and then they can decide if they want to pay the CPM rate,right?

And, you know, that's where you need to then get new payment rails for the internet that can do, you know, millions of transactions a second for all these little micro payments. All these things that we've been talking about for years, you know, micro payments, you know, always are two years away.

And, you know, I think agents will actually push that forward and actually I think we will see a world where actually there is a new economic model for content creators.

**Sarah Guo** [16:02]
Who do you think the players are or what are the variations of that infrastructure,right? Like I think obviously Cloudflare has an important role to play here, but the rise of personal agents like Instinct and Muse have also made this like an immediate question from a transactional perspective.

**Dane Knecht** [16:21]
I don't think everyone knows on the other side what that business outcome is going to look like. You know, everyone's been through these transformations where the internet, you know, it's always going to kill Main Street. It didn't kill Main Street.

Their small businesses are still thriving. But, you know, there were people that were displaced. And so there's always kind of some skepticism there. And so, you know, there's some natural inclination to like just block them out,right? You know, you saw Amazon just say, "No, no personal shopping there from an agent."

I mean, I think these are all people that are, you know, that messy middle where they're trying to figure out what that path forward is and trying to keep some of their negotiation leverage there. I think we will get to a place where

Muse and Dots and all those personal agents will be able to accomplish tasks seamlessly as you as a human could do, even easier. But I don't know if the technology is quite caught up to that yet and the business models either.

**Sarah Guo** [17:24]
Is there anything you're personally excited about on the roadmap or just from a technical exploration perspective on the Cloudflare front that you like feel like is going to enable this ecosystem in the future?

### Monetization

**Dane Knecht** [17:38]
Well, yeah, I mean, absolutely. On the content side, you know, we've released a conference monetization gateway where we make it easy for people to put their sites behind us and take micro transactions for their content,

including providing wallets for agents to use

and a stablecoin that's going to be able to help kind of facilitate that transaction as well. You know, we're also doing things with OpenAI, a project called Signals where we're making it easier for agents to actually discover content instead of having to crawl the entire internet and continually go back to pages.

We're making it easier for them to know what's new content and actually kind of more proactively do that. So, you know, it's a win for the AI labs as well as a win for the site owners not having to get constantly crawled and pay for that infrastructure.

So, you know, there's a whole set of kind of new laws on the internet that I think really excited to see those kind of grow to see, you know, the, you know, X402 from when Stripe came out with it.

I loved it because it went back to the 402 HTTP code, which was, you know, Tim Berners-Lee created 20 plus years ago, 30 plus years ago for payments. So like, you know, kind of always knew that you needed a payments, a better payments rail on the internet, but we haven't been able to, you know, kind of figure it out.

And I think agents will lead us to that as well.

**Sarah Guo** [19:10]
When you think about these initial products that you've released, like is there, what would success for a content creator or somebody who ostensibly had a website on the internet that they were monetizing with ads before? Like what's the killer success case you guys need to see or show the world where it just becomes like a default to do this?

**Dane Knecht** [19:34]
Well, I mean, I think where instead of them paying us, we're paying them,right? You know, where it's covering their infrastructure costs, it's providing money

that supports their livelihood. You know, if we can actually increase the number of content creators out there, which, you know, the last generation of the internet has done very well, like with YouTube and TikTok and X, I think we can actually make it even exponentially larger.

**Sarah Guo** [20:04]
And when you look at the Workspace OS, like release and ecosystem work that you guys have been doing,

### Cloudflare OS

**Sarah Guo** [20:14]
what is, what are you hoping are the like biggest use cases of it coming up?

**Dane Knecht** [20:18]
This is transitioning a little to a new release we called Cloudflare OS. And Cloudflare OS is our internal platform for how our employees use AI every day where we, you know, and it actually goes back to security roots where, you know, we were very worried about customer data, internal data leaking out.

And we wanted to make sure that our employees could, you know, use all the new AI tools and models, but have a way to safely deploy them internally to understand how that data is being used and flowing and have kind of good governance over that.

In addition, the reason my company called Cloudflare OS is we started to realize that, you know, as we have more and more agents, you know, we need to be able to teach them what makes Cloudflare special, how we build at Cloudflare.

You know, and for humans, we did that through Cloudflare University and the onboarding process and, you know, meeting your colleagues and working with them. But we needed to kind of more clearly define that and document that for agents.

And so Cloudflare OS is both the platform where our employees can build anything that they want in a secure way without having to, you know, security is not something that they have to think about. You know, the BR that's trying to create a call list, they don't have to worry about security.

It's just built in there for them. But they also know that it's being done in a way that is consistent with the themes and values of how Cloudflare works.

**Sarah Guo** [21:43]
And why release? I think of this as like somewhat different from the other products that Cloudflare offers today in that it is not infrastructure. It's like an end user facing product for managing AI use within the organization. Like why release it?

Or do you think of it as a departure?

**Dane Knecht** [22:01]
Yeah. I mean,

for one, it was, it's kind of a reference architecture really. It's, you know, over the past 10 years, we've built some our developer platform where we started with just a simple serverless compute primitive and now, you know, over a hundred different developer platform tools from storage to databases to durable objects.

We wanted to show someone, show how you can build it completely on Cloudflare and how we build things internally and kind of

give people the option to, you know, fork it and build their own version of their own company OS on top of it.

**Sarah Guo** [22:41]
And what is a way in which it is helping your people and your organization that you think is like clear success?

**Dane Knecht** [22:49]
It's actually used more by our go-to-market teams and our G&A teams than it is even used by our engineering teams. You know, it's, you know, giving them a very fast, you know, early on I thought, you know, we saw training was the biggest gap for people using AI.

And when you just kind of point them at a chat interface or

another vibe putting tool, it's too much of a blank slate for them. And so, you know, this is a platform that's designed, you know, you design everything for it, for how your business works, and they immediately can go there and, you know, get access to all the data that they need.

They can get guided and taken through where they quickly realize, oh, this is not, you know, trying to replace me, this is trying to make me better at my job. And that's been really exciting to see just all the different ways people at the company have done that.

You know, a lot of the, how this conference got put together, the marketing team, I mean, built hundreds of internal tools where previously, you know, they would have to go out and work with different agencies, but they can, you know, the speed that they can move with a smaller team now is just amazing.

**Sarah Guo** [24:01]
Assuming this is a continuation of, you know, starting in May, you guys did this restructuring to be an like sort of AI-first organization. What else, reflecting on, you know, the last number of months has worked or not worked that you would recommend to others?

### Restructuring

**Dane Knecht** [24:19]
Yeah. I mean, so, you know, the restructuring was hard. You know, saying goodbye to some of our colleagues was the first time Cloudflare ever did it. You know, we had always been very careful not to over-hire. And

this was for different reasons. This is because we saw the way AI was transforming some teams and how they work. And, you know, we kind of thought of it as like a factory floor where you needed to actually rebuild, rethink how things happen.

You know, there were some processes or some, you know, the way a product manager would send an email forward to customers, it involved three teams. You know, it involved a team where, you know, you went to BI to go get a cohort of the people that you wanted to send an email to.

You went to the marketing team to help you design the email and get the voiceright. You went to the marketing operations team to get the email out where, you know, you're never going to be able to actually redesign that process unless you kind of make some hard changes and force people to kind of rethink, well, what actually is theright way to do that?

And it might not involve those three teams at all. It might just involve a new way of doing things. And, you know, so what came out of that was a tool where a product manager now goes in and describes the cohort.

They write an email, it rewrites it in Cloudflare's voice, it sends it to their manager, and the manager can click approve if they like it, and it goes out where, you know, we didn't go from three teams getting faster at doing that.

We went to, you know, one tool that can kind of let a product manager now have a superpower that they didn't have before.

**Sarah Guo** [25:56]
When you think of all of the different workflows and tasks within an organization that's got about 5,000 people like Cloudflare, it's like, how do you decide who the owner is of rethinking that process? And like, was it bottoms up from one of those teams in this case?

Like, is it an organizational leader? Like when you have so many of these things that could change to be faster and make more sense as you describe?

**Dane Knecht** [26:20]
Yeah. I mean, Cloudflare has always had a thousand more things that we could ever do, whether it was, you know, products and features or things like this. And so we always look for people that are going to take initiative and ownership.

And so, you know, it doesn't really matter if you're ordering it, you know, one to a hundred and try and prioritize it. Any one of those hundred things are probably safe if someone's passionate about it. So, you know, we kind of gave people at the beginning of last year the tools and started seeing how they were experimenting.

And, you know, people quickly kind of broke out into different cohorts of people that were really leaning in and, you know, just naturally doing those things and taking ownership of them. So, you know, I think it was very bottoms up from the how we're going to do those things.

**Sarah Guo** [27:06]
It's been a major topic of conversation for organizations as to how they should think about token budgeting,right? Like, you know, what's a reasonable amount to spend? How do I think about this versus human capital? Who should get tokens even within our organization?

### Token budgeting

**Sarah Guo** [27:23]
Like, what is your philosophy on this now?

**Dane Knecht** [27:25]
It's definitely changed. You know, and each model kind of makes you rethink it as the cost dynamics change. But, you know, in general, you know, I think it's kind of silly that we're actually asking humans to decide which model to do each task.

**Sarah Guo** [27:42]
Like it should be intelligently routed.

**Dane Knecht** [27:43]
Yeah.

**Sarah Guo** [27:43]
Okay.

**Dane Knecht** [27:44]
Exactly. Like humans just aren't going to ever be good at that,right? Like, you know, but, you know, early on you do it kind of brute force and you get people to experiment and you say, hey, well, you know, let's try for the planning part.

Let's use the top frontier models for that. And then, you know, once the plan's written out and agreed upon and iterated on, then, you know, potentially some more of the open source models can go do that. And then you go back and you have one of the frontier models do the reviews and the red teaming and things like that.

Now, all that today is, you know, fairly manual. But I think, you know, as we start to learn and experiment, I think we will become intelligent. And, you know,

and, you know, over time there'll be more governance on that. You know, Cloudflare has been in a place where, you know, largely we've let people use the models and tools that they see fit for the job. You know, we don't have a token leaderboard.

We don't encourage it. We encourage people to focus on the output. The, you know, they're not graded on, you know, if they used it, they're graded on, you know, what they've shipped and what they've built. But then, you know, we do have some kind of, we found like kind of like giving people notices.

Like, you know, if you happen to use $300 a day, they'll just get an email,right? Like it says, hey, did you know? And because most of the time people didn't realize they were doing it. And they're like, oh, huh.

And then I think it gets to like 500, it goes to your manager, if it goes to a thousand, it goes to a VP in a day. But it's, you know, those things rarely happen. And usually it's, people are actually happy to get those notices.

**Sarah Guo** [29:26]
Where do you think you are in your overall evolution of like take things just for the current level of capabilities of models, assume no new development of, you know, rethinking the way different things at Cloudflare are done from go-to-market to engineering to be?

### Intern strategy

**Dane Knecht** [29:41]
Oh, I think we're still so early. You know, Cloudflare did something, you know, two years ago we had about 200 interns. And what we saw was those interns, I think were, you know, really the first, you know, fully kind of AI native, you know, where if you took AI away from them, I don't think they would be able to work.

**Sarah Guo** [30:02]
Oh no. Is that good?

**Dane Knecht** [30:05]
Actually, I think it is. I think they know how to use the tools better than us. It's just like you took Google away from me, you know, over the last 10 years, you know, like, you know, how would I learn anything,right?

Like that's where they reached to first. And we actually saw the managers that had interns from that summer actually adopt AI faster. So then this past summer, this past year, we decided we were going to hire a thousand interns.

So that's, you know.

**Sarah Guo** [30:31]
For 5,000 person company.

**Dane Knecht** [30:32]
Yes.

**Sarah Guo** [30:32]
Wow.

**Dane Knecht** [30:33]
Yeah. Over the course of the year.

**Sarah Guo** [30:35]
Yeah, still.

**Dane Knecht** [30:36]
Yeah. And it's just been amazing because, you know, one, just the number of ideas that come from the interns, you know, that's something that, you know, with AI, lets people experiment faster. And so having more ideas, you get more people in there.

But also just seeing the way that they use AI, like, you know,

if they need to fill out a form, they reach for AI. And I still find myself, you know, there's so many things where I'm like, oh, wait, I've just been doing this manually still. And then I'm like, yeah, oh, wait, actually I should be using AI.

And I don't think that that's

natural for most of us yet. And I think that will take time. And so, yeah, I think we're still very early in our AI journey.

**Sarah Guo** [31:20]
I love it though. The mass intern strategy. I run a very small partnership and we always have a, you know, an intern or two floating around a team of about a dozen for, you know, ideas and intern energy and all that.

But we just converted one of our, you know, in-college interns to full-time and JED and had him drop out and everything. And I think that you'reright where maybe like the dial should be turned up on the amount of DNA that is really, really native to these tools.

**Dane Knecht** [31:52]
Yeah. And I think it's great for them as well too. I mean, you see these schools like Waterloo and Northeastern, like with these co-op programs. And, you know, from what we've seen is people that get that type of real work experience just, you know, are some of our best employees many years in and go on to do, you know, great things.

**Sarah Guo** [32:14]
It is interesting or encouraging though that instead of it being like, hey, here's a master plan, we're going to map out every department and subdepartment at Cloudflare and be like, here are the key workflows. We're going to try to re-engineer the workflows.

It is instead a strategy of like, assume people are rational about their tokens if given information about their usage, hire a bunch of interns and assume high agency people we have here are going to go like, you know, shift those workflows to something more intelligent over time.

**Dane Knecht** [32:42]
Yeah. And then brute force and just try things. You know, early on we had an email alias called agents@cloudflare where we, this is back in January of last, of this year, where we just told people to, you know, if you're curious, can I use AI for this?

You know, email in, your jobs to be done. You know, and that email list would reply back to the skill file.

And for a while people thought it was actually, you know, thought it was all AI and automated. It was actually a bunch of interns that were, you know, helping people write skill files. But that let us kind of get to see, you know, in a way that's not going out and like, you know, hiring a consulting firm to come out and like catalog every single thing to be done.

It's, you know, people that are recognizing, hey, this seems harder than it needs to be. I wonder if an agent can help me here.

### Hiring & curiosity

**Sarah Guo** [33:34]
Is there anything else about the last year that's changed beyond the intern level, like your hiring philosophy, what you looked for in people to bring on?

**Dane Knecht** [33:45]
Well, I mean, one of the biggest things we always looked for is curiosity. And so, you know, if people aren't using AI at home in their personal life and all that, I mean, that's a pretty big red flag.

Like, I mean, just how can you not be embracing AI if you have, you know, a level of curiosity out there? So, you know, I think just doubling down on some of those kind of key characteristics of, you know, drive and curiosity and empathy

play even better for this next set of people.

**Sarah Guo** [34:20]
Well, I mean, I'm an alpha user of all products and technology forever. So, but I think the answer is like fear and skepticism,right? And so I guess implicit in this claim about like curiosity, including in experimentation with frontier tools is like that needs to, that needs to be a core principle for people who work at Cloudflare.

**Dane Knecht** [34:47]
Yeah. I mean, a lot of things that we're doing, we're not, you know,

we're inventing new ways of doing things. I mean, the scale that we operate at for the number of sites behind us and the amount of free services we give away, it's just not possible with off-the-shelf technology. So you need people that can come in and literally invent new things.

And that, I mean, that starts with curiosity.

**Sarah Guo** [35:11]
I have a couple of questions for you to close out and mainly about security, both and how to do that in a new way, both at Cloudflare and then how you think it's going to work in the broader ecosystem.

### Red teams

**Sarah Guo** [35:23]
What do you think is the like process by which if half the engineering team is going to go work on security and just like fix 70 years of vulnerability, how do you attack that problem? Internally, assuming that not all software is perfect, even amazing production internet scale software.

And then, you know, we'll talk about the ecosystem.

**Dane Knecht** [35:42]
Yeah. I mean, I think largely, you know, we're, you know, today we have one red team. I think by the end of the year next year, we will probably have four or five red teams. So, you know, break people out into groups that are actually, you know, being more adversarial and actually like going out and actively trying to find vulnerabilities.

**Sarah Guo** [36:04]
Why is that the approach when instead of like, oh, we're finally going to get like our patching systems,right? Or our processes,right?

**Dane Knecht** [36:12]
I mean, processes and those things could always be better. We have a vulnerable process. We have a patching process. Are they perfect? No. But, you know, the best way I found for those sorts of things to get better is to, out of crisis kind of comes,

that's how you get those things fixed. You know, when the vulnerabilities get too big for the engineering team to just do them using their normal processes, you know, engineers are inherently lazy in a good way. And so they will automate away those things as they need to, as they kind of realize that there's no other way to do that.

So, you know, I think the focus needs to be on the discovery side. And, you know, people will do theright thing and figure out how to patch it in a scalable way over time.

**Sarah Guo** [37:02]
Do you express a philosophy on just prioritization of this versus like shipping new features? Because I guess my understanding of, it's not as if we didn't have a backlog of like vulnerabilities in every large piece of software to go fix.

It was just like whether or not that was a priority for engineering teams previously as one common state.

**Dane Knecht** [37:25]
I mean, I think you need to figure out how to do both. You know, it might come at the expense of cost optimization or other things, but I mean, a company's, if you're not shipping, you're dying. It's how we think about it.

So.

**Sarah Guo** [37:40]
So it's not going to be a freeze in order to do this.

**Dane Knecht** [37:42]
Oh, no, absolutely not. You know, you need to be, I mean, you need to be able to continue building and shipping new things, but you need to be able to make sure that you're doing it, you know, securely and patching and fixing bugs.

And

I think if we do itright, we'll actually be able to do them both faster with AI. I mean, while AI has become the

tool that's causing this kind of security apocalypse, vulnerable apocalypse to happen, it is also the tool that can help us solve it faster as well,right? So it's on both sides of it.

**Sarah Guo** [38:20]
Oh, where do you fall in this debate of the role, like good and bad of open and closed models on the cyber front,right? Because the argument, the broad argument is, well, having defensive and offensive cyber capabilities, you know, they're parallel in the same model despite safety and guardrails being built into them.

And so you can't have open defensive models without them also being offensive. And then the argument sort of, or the concern around only having security capabilities in a set of frontier closed models is, well, how are you going to release them?

And you can't possibly get them out to the distributed set of actors that need to defend. Like what do you think? Or did I characterize that correctly?

**Dane Knecht** [39:09]
Yeah. I mean, I think over time it all has to be open because everyone's going to get access to it anyways. And, you know, the.

**Sarah Guo** [39:17]
That's how open source tends to work.

**Dane Knecht** [39:19]
Yeah. The bad guys, the bad guys are going to find a way to get it. So all you are, and the, you know, the top companies will have a way of getting it as well. So I think you're just excluding a mass population that just doesn't have the access that probably needs it.

**Sarah Guo** [39:37]
And then so, final question, just because, you know, Cloudflare, very mission-oriented, does carry the weight of some, you know, internet resilience on its back. What about all of the people who can't afford to have four red teams or like never had a red team to begin with that run important systems?

Like how do they protect themselves?

**Dane Knecht** [39:58]
Also, you know, I would say put your sites behind Cloudflare, put your infrastructure behind Cloudflare, build on Cloudflare. You know, when these major vulnerabilities come out, like the React one last year, we don't just give it to our paid customers who have like the advanced firewall and WAF capabilities.

We give it to everyone. We view those kind of things as part of the utility that we offer. And, you know,

we want to provide, I just like, you know, we want to provide performance for everyone. You know, before Cloudflare started, you know, there were probably 3,000 sites in the world that had a large scale CDN to cache their stuff at the edge.

You know, now there's hundreds of millions of sites that do. So, you know, Cloudflare has always viewed our goal as providing, democratizing the security and performance for everyone.

**Sarah Guo** [40:48]
Awesome. Thanks so much for doing this, Dane. And you guys heard it first. We have the agent kitties and the vulnapocalypse. Am I pronouncing thatright?

**Dane Knecht** [40:58]
Yeah.

**Sarah Guo** [40:58]
Yeah. But it's resolvable. Allright. Thanks so much, guys.

**Dane Knecht** [41:01]
Thanks.

**Sarah Guo** [41:06]
Find us on Twitter@NoPriorsPod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.

---

This library is powered by PodHood (https://podhood.com), the podcast website platform.
