Cloudflare CTO Dane Knecht on AI as both threat and remedy for internet security

The host begins by observing that AI has become the tool causing a kind of security apocalypse, and that it is also the tool that can help solve the problem faster. He notes that the conversation has moved from how to get everyone using AI to a race to define what AI security means.

Knecht says that over the past year his priorities have constantly shifted, but they still trace back to Cloudflare's mission of helping build a better internet, which he says has not changed in 15 years. He describes the goals as keeping the internet secure and letting everyone get content online easily and scale it.

Knecht describes a case that he says is coming out and being discussed publicly: one of the keys protecting all TLS and WebPKI, which secures connections across the internet, was reportedly reverse-engineered because its seed was based on a single number, four. He explains that cryptographic security depends on randomness, which is why systems use sources such as lava lamps for good random numbers.

He argues that AI can perform this kind of work continuously. Unlike a single engineer doing red-team work, AI does not sleep. He also says that last week OpenAI reportedly found the seeds for the NIST curves, which were released back in 2000. In his view, the models are improving quickly, and the progress now comes not only from better models but also from harnesses and from how people use frontier models to find exploits.

Asked about the humans behind these attacks, Knecht begins to explain that some of the larger vulnerabilities that have surfaced recently did not come from nation-state actors.

The host reports hearing serious technical people say that everything important should be air-gapped, and that the internet may not be trustworthy for a while, with critical infrastructure and government services assumed to fail at some point. He asks whether that is Knecht's expectation.

Knecht says he is less doomsday. He argues that humans are adaptable and can move fast when needed, and that he has seen the industry partner together during such events, with usual competitors working to protect critical infrastructure. He also says that the move over the past five to ten years from castle-and-moat security toward zero trust puts many companies in a better position, because devices and data flows are secured in different ways rather than relying on a single fortified perimeter.

The host recalls that Greg Brockman of OpenAI, on the No Priors podcast the previous week, said he thought that by 2027 people would spend as much money on tokens for security as for engineering. He asks what will drive that shift.

Knecht says there is an emergency, so security spend will move forward over the next 18 months. That spending covers both the red-team side, which actively tries to exploit systems, and the defensive side, which prepares systems. He does not expect this to last forever. He anticipates that after about 18 months, as resources refocus, new software productivity will rise sharply, though he says there is a period in which security must be on everyone's mind.

The host raises the issue that today's models do not naturally aim to reduce surface area. He describes asking a model to do something and then asking it to try again with a tenth of the complexity. He asks whether this is a role for engineering and security engineering, or something the models will fix.

Knecht says the problem has to be addressed at all levels: engineers who write code, including product engineers, need training, and security teams must remain involved to provide checks, balances and a second set of eyes. He describes processes Cloudflare has defined internally, including a document it calls codex, built from RFCs on how to build secure and available software. That guidance is condensed into ten principles agents can follow during review. He describes agents handling much of the routine output while people run performance and security tests, and says that, as before, adversarial reviews are needed.

The host notes that agent-driven traffic across the internet has outpaced most people's expectations and asks how the web may look a year from now.

Knecht describes the present as a messy middle, where humans and agents share the same internet rails. Agents parse a web designed for humans, which he calls inefficient. A human can visit a site, read it, see the ads and take part in the economic model, while an agent can visit a thousand pages in the same time. He says this is great, but a new economic model is needed to support it.

Asked whether there can be parallel webs, Knecht says he thinks the content will be delivered in ways that meet agents and humans where they are, rather than in the same place. He criticizes current translation layers that help agents read web pages as extra token burn that probably should not be necessary, noting that APIs have long let computers talk to each other but covered only a small fraction of the information on the internet. He says agents want to explore all of it, so the cost for site owners must come down and a way to compensate them must be found.

The host raises the rise of personal agents as making the transactional question immediate. Knecht says he does not think everyone knows what the business outcome will look like on the other side. He recalls that people said the internet would kill Main Street, but it did not, and that small businesses are still thriving. He acknowledges that some people were displaced, so there is always skepticism.

He notes that some companies are blocking agents, citing Amazon's refusal of personal shopping agents, and describes these as parties trying to keep negotiating leverage during a messy transition. He expects agents such as Muse and others to eventually complete tasks as seamlessly as a human, but says neither the technology nor the business models are quite there yet.

Asked for the killer success case, Knecht says the idea is that Cloudflare pays site owners rather than them paying Cloudflare. He says this covers their infrastructure costs and provides money that supports their livelihood. He argues this could increase the number of content creators, saying the last generation of the internet did well with YouTube, TikTok and X, and that the new model could make the creator base even larger.

Cloudflare OS is Cloudflare's internal platform for how employees use AI every day, and it grew from security concerns about customer and internal data leaking out. Knecht says the company wanted employees to use new AI tools while having a safe way to deploy them internally, understand how data flows, and maintain governance.

Knecht describes it as a reference architecture. Over the past ten years, Cloudflare has built a developer platform starting from a serverless compute primitive, and it now has more than 100 developer platform tools, from storage to databases to durable objects. The company wanted to show how someone could build an entire system on Cloudflare and gave others the option to fork it and build their own company OS.

Knecht says the restructuring was hard, and saying goodbye to some colleagues was the first time Cloudflare had done that. He explains that the company had always been careful not to overhire, but this time it changed because it saw how AI was transforming some teams and how they work. He describes the company as a factory floor that needed to rebuild how things happen.

He gives an example: a product manager sending an email to customers used to involve three teams, one to pull a cohort from BI, one in marketing to design the email and get the voice right, and one in marketing operations to send it. Knecht says the process would not be redesigned without hard changes. The result was a tool where a product manager describes the cohort, writes the email, has it rewritten in Cloudflare's voice, and sends it to their manager for approval, with the email escalating up the chain only in rare cases.

Knecht says the marketing team built hundreds of internal tools that previously would have required outside agencies, and that the speed a smaller team now has is remarkable. He notes that the go-to-market and G&A teams use the platform more than engineering does.

The host asks how an owner is chosen for rethinking processes across an organization of about 5,000 people, and whether the work is bottom-up or led by organizational leaders. Knecht says Cloudflare has always had far more ideas than it could ever pursue, so the company looks for people who will take initiative and ownership. He says that it does not much matter how items are ranked from one to a hundred, since any one of those things is probably worth doing if someone is passionate about it.

He says that early last year the company gave people AI tools and watched them experiment. People quickly formed cohorts that leaned in and took ownership of the work naturally, so he describes the process as very bottom-up.

The conversation then moves to token budgeting. Knecht says the philosophy has changed as cost dynamics shift with each model. He says it is silly that humans are asked to decide which model to use for each task, arguing that tasks should be routed intelligently.

Asked how far Cloudflare has evolved, Knecht says the company is still very early in its AI journey. Two years ago it had about 200 interns, who he says were the first truly AI-native group; without AI, he doubts they could work. He says they know how to use the tools better than the company does, comparing it to how he would rely on Google. He adds that managers who had interns that summer adopted AI faster, so this past year Cloudflare decided to hire a thousand interns for a company of 5,000 people.

Knecht says the interns have produced many ideas, since with AI less people can experiment faster. He notes that they reach for AI even for filling out a form. He admits that he still catches himself doing tasks manually that he could do with AI, and says that habit will take time to change for most people.

The host says he loves the mass intern strategy. He runs a small partnership with a team of about a dozen and often has an intern or two, and recently converted one college intern to full time. He suggests that the dial should be turned up on how native to these tools people are.

Knecht agrees that such programs are good for interns, noting that co-op programs at schools such as Waterloo and Northeastern produce people who get real work experience and often become some of Cloudflare's best employees years later. He contrasts this with a master plan that maps every department and workflow. Instead, Cloudflare's approach assumes people will act rationally about their tokens if given information about their usage, hires many interns, and trusts high-agency people to shift workflows over time.

Knecht describes early experiments, including an email alias called agents at Cloudflare that, starting in January of the current year, invited people to ask whether AI could help with a job to be done. Those requests were answered and supported with skills.

Knecht says one of the biggest things Cloudflare has always looked for is curiosity. If people are not using AI at home in their personal lives, he says, that is a pretty big red flag, because it is hard to see how someone curious would not be embracing AI. He argues that drive, curiosity and empathy matter even more for the next set of hires.

The host, who describes himself as an alpha user of all products, suggests that the reason some people do not use AI is fear and skepticism. Knecht says that curiosity, including experimentation with frontier tools, needs to be a core principle for people who work at Cloudflare. He explains that the company invents new ways of doing things because of the scale it operates at, the number of sites behind it and the amount of free services it gives away, which is not possible with off-the-shelf technology. He says such people must be able to invent new things, and that starts with curiosity.

The host asks how security work should be prioritized against shipping new features, given that every large piece of software has long had a backlog of vulnerabilities and it was up to engineering teams to make fixes a priority. Knecht says companies need to do both, which might come at the expense of cost optimization. He says that at Cloudflare, if you are not shipping, you are dying.

Asked whether there will be a feature freeze, Knecht says absolutely not. Teams need to keep building and shipping new things while making sure they do so securely and patching and fixing bugs. He says that if done right, both can be done faster with AI, and repeats that AI has become the tool causing this security apocalypse while also being the tool that can help solve it faster.

The host then raises the debate over open and closed models on the cyber front, noting the argument that defensive and offensive cyber capabilities sit in the same models despite guardrails. Knecht's answer is not fully developed in this passage.

The host asks about organizations that cannot afford multiple red teams, or never had one, yet run important systems. Knecht advises putting sites and infrastructure behind Cloudflare and building on it. He points to major vulnerabilities such as last year's React issue, which Cloudflare did not offer only to paid customers with advanced firewall and WAF capabilities. It gave the fix to everyone, because it views such protections as part of the utility it offers.

Knecht compares this with the past: before Cloudflare, there were perhaps 3,000 sites in the world with large-scale CDN caching at the edge, and now hundreds of millions of sites do this. He says Cloudflare has always viewed its goal as democratizing security and performance for everyone.

The host closes by thanking Knecht and directing listeners to the show's platforms.