No Priors

Dane Knecht on AI as both the security apocalypse and the cure

Dane Knecht· CTO at Cloudflare
·~41 min·English·No Priors
AI SafetyAgentsAI InfrastructureBusiness Strategy
TL;DR

Cloudflare CTO Dane Knecht on why AI has set off a weekly drumbeat of serious vulnerabilities, why the dangerous attacker is now an amateur with an agent rather than a nation-state, and how he is rebuilding both Cloudflare's defenses and its own org around AI.

01Core Framing

The security apocalypse is also the cure

AI is driving an explosion of discovered vulnerabilities — Log4j-scale incidents that used to surface about once a year now surface weekly — but Knecht's framing is that the same capability is also the fastest way to find and fix them.

AI has become the tool that's causing this kind of security apocalypse is also the tool that can help us to solve it faster as well

— Dane Knecht, No Priors
Key Insight
The symmetric framing hides a timing asymmetry: attackers get the speed-up first, because finding one hole is cheaper than proving all of them closed. That is why Knecht expects a painful transition — more than half of engineering time spent on security — before software ends up more secure than it was.
02The New Attacker

Agent kitties

The biggest recent vulnerabilities, Knecht says, were not found by nation-states but by high-schoolers and bug-bounty hunters pointing agents at targets — the 'agent kitties' the episode names, running what used to be capture-the-flag contests against real production systems.

it's given the tools that honestly the nation states have had in some forms for much longer

— Dane Knecht, No Priors
Key Insight
When tooling that was a nation-state's moat becomes a weekend project, the scarce resource stops being capability and becomes attention. The threat model shifts from who is sophisticated to what is exposed — there are simply more eyes on more targets, and the financial bounty keeps them looking.
03The Measured Response

A little less doomsday

Asked whether we should just air-gap everything and assume critical infrastructure falls, Knecht is 'a little less doomsday': a decade of moving from castle-and-moat security to zero trust already shrank the blast radius of a breach, and in a real crisis the usual competitors partner to protect shared infrastructure.

people that you usually think of competitors are no longer competitors. we're just literally protecting critical infrastructure

— Dane Knecht, No Priors
Key Insight
Knecht's calm rests on a migration that already happened: a decade of zero-trust work means the AI shock lands on a system that was already being hardened, not on a single flat wall. His answer to the air-gap question is to tighten what exists, not to disconnect.
04Defense, First Principles

Shrink the attack surface

The highest-leverage defense is also the oldest — reduce the attack surface by writing and running less code — which is exactly what today's verbose models are bad at, so human judgment still has to set the design.

looking at all your software and saying how can I simplify this

— Dane Knecht, No Priors
Key Insight
Reducing attack surface is the one defense that compounds: every line you never write is a bug you never have to find or patch. But it runs against how current models behave — they add complexity by default — so the near-term bottleneck is not detection, it is getting AI to write less, not more.
05The Agentic Web

The messy middle

We are in a 'messy middle' where humans and agents ride the same rails: a web built for a human who reads one page and sees an ad now also serves agents that hit a thousand pages in the same time and see none of them.

we're in the messy middle right now where humans and agents are

— Dane Knecht, No Priors
Key Insight
The mismatch is not technical, it is economic: the web's business model assumes a human who can be shown an ad, and an agent quietly breaks that assumption at a thousand-to-one scale. That is why publishers' first instinct is to block agents rather than serve them.
06The Economic Model

New rails for the web

Knecht's fix is a new payment rail — pay-per-crawl, wallets for agents, and the long-dormant HTTP 402 status — so an agent can decide, page by page, whether to pay the content's going rate instead of scraping it for free.

that's where you need to then get new payment rails for the internet

— Dane Knecht, No Priors
Key Insight
HTTP 402 has sat unused for decades because there was never enough volume to justify the plumbing. Agents supply that volume — so the thing that finally makes micropayments real is not a better payment technology but a buyer that reads a thousand pages in the time a human reads one and can judge, per page, whether the content is worth its price.
07Secure by Process

Ten commandments for the agents

Cloudflare writes its security principles down, boils them into about ten rules its agents follow at every review step, and keeps humans and frontier models doing the adversarial checks deeper in the 'factory.'

it kind of gets boiled up into some ten commandments

— Dane Knecht, No Priors
Key Insight
Writing the rules down is a bet that security can be a property of the process rather than of any one engineer. If the rules live in the review step, a mediocre agent and a careful human are held to the same bar — which is how a 5,000-person company multiplies its output without lowering its standard.
08AI-Native Org

The thousand-intern bet

Cloudflare's AI-native strategy is less a master plan than a bet on people: hire a thousand interns into a 5,000-person company, give everyone tokens and usage visibility, and let high-agency employees rebuild their own workflows.

we decided we're going to hire a thousand interns

— Dane Knecht, No Priors
Key Insight
Hiring a thousand interns is less a labor decision than a culture one. The fastest way to make a workforce AI-native is to flood it with people who have never worked any other way — Knecht noticed the managers who sat next to AI-native interns adopted the tools faster themselves.