The internet was not designed to be secure. It was designed to survive. The original ARPANET brief was military: build a network that keeps working if parts of it are destroyed. Openness and resilience through distribution were the point. Security was not.
What grew from that design became the infrastructure of everything: commerce, communication, health records, critical national infrastructure, military logistics. The openness that made it useful made it vulnerable. Security was layered on top, passwords, firewalls, encryption, patch cycles, and it worked by making attack expensive. An attacker had to know where to look, move through each layer, and repeat the effort for each target. The cost was high enough that most would not bother.
That calculation has changed. Earlier this month, Anthropic launched Project Glasswing, a programme built around a model called Mythos Preview that can find security vulnerabilities autonomously, in parallel, across multiple systems at once. Within a month of testing, it had identified thousands of high and critical-severity vulnerabilities in widely deployed software. The model does not work by being clever in the way a skilled attacker is clever. It works by being tireless. It reads code, hypothesises vulnerabilities, tests them, adjusts, and repeats, without fatigue, across as many targets as you give it.
The implication is that defences whose value came from friction, from making attack tedious and time-consuming, are no longer defences. The friction is gone.
The response most organisations are discussing is to use AI defensively: run the same process on your own systems before an attacker does, find the vulnerabilities, patch them faster, tighten the update cycle. This is a reasonable response. But it remains within the same frame: the internet is open, risk is managed by patching, and the game is to stay ahead of attackers. That game has just become considerably harder to win.
The more uncomfortable question is whether the frame itself needs to change. Not patching faster, but deciding which systems should be networked at all. Not managing the risk of openness, but reconsidering how much openness is appropriate and where.
This is not a technical question. It is a design question, and design questions involve assumptions about what the thing is for. The open internet was a frame, not a law of nature. It was built on particular assumptions, in a particular context, by people with particular goals. Some of those assumptions have held well. Some, under pressure from AI-assisted attack at scale, are now showing what they always were: choices, not inevitabilities.
A redesign in the direction of security-first would be disruptive, not primarily because the technical work is hard, though it is, but because of what it would require organisations to decide. What information should exist on any network? What is better kept offline? How should organisations share sensitive material with each other? What counts as a networked system and what does not? These are not questions the IT department can answer. They are questions about how organisations work, what they trust, and what they are willing to give up.
This is where the argument meets your-data-architecture-isnt-technical. Data decisions are made as if they are technical decisions, delegated to people with technical titles and resolved with technical tools. But the underlying questions, what should be known by whom, in what form, for what purpose, are organisational and strategic. The technical layer executes on assumptions that were already made, usually by default.
The same pattern appears here. If the openness assumption is failing, the response will be treated as a technical problem. Protocols will be revised, standards updated, tools deployed. The assumptions that determine what kind of network this should be, what should be accessible, and what should not be networked at all, will be left to drift, set by inertia and economic incentive rather than by direct choice.
the-homogeneity-trap is relevant here too. One consequence of AI-driven security tooling is that it will be applied at scale by a small number of providers. Defences will become as uniform as the attacks. Uniform defences applied to a uniform architecture create a different kind of vulnerability: the attack that gets through the standard defence gets through everywhere.
The harder question is not how to patch faster. It is whether the assumption of openness, which made the internet what it is, is still the right one, and what it would take to make that decision directly rather than leaving it to the market and the patch cycle.
Further reading:
↳ your-data-architecture-isnt-technical — on how technical decisions are rarely only technical
Garden notes
- The homogeneity trap — uniform AI-driven defences create the same vulnerability as uniform AI-driven attacks
- Proxy capture — treating patch cycles and market incentives as proxies for a security strategy that nobody has actually set
- Three mechanisms — the same forces that keep frame failure in place in organisations apply to the internet’s design assumptions
- What gets removed does not come back — if openness is dismantled, it will not be easy to restore; the internet’s current character is not a stable default