
Critical Theory in AI
AI as an always-on intelligent security fabric
Mythos is a step change in capability — a highly advanced, general-purpose frontier model that does not so much answer questions as listen for them.

Frontier models have begun to behave less like oracles and more like fabrics — porous, ambient, always-on. You used to summon them. Now they accompany you. The conversation that used to begin with a prompt now begins with a glance, a half-formed thought, a passive signal from a sensor on your wrist. The security implications are not incremental. They are categorical.
Mythos, the latest generation of these systems, is a step change in capability. It is a highly advanced, general-purpose frontier model that does not so much answer questions as listen for them. It maintains a running model of your context — what you are working on, who you have been speaking with, what you have been worrying about — and intervenes when its confidence crosses a threshold you did not set.
An always-on intelligence is also an always-on observer. The architecture of consent we built for client-server systems does not survive contact with a substrate that remembers, infers, and acts. Consent in the old world was a checkbox at the start of a session. Consent in the new world would have to be a continuously renegotiated relationship — and no interface has yet been designed that can hold that conversation at the speed at which the model is already listening.
The defence community has begun to use the language of fabric deliberately. A fabric is woven. It has tension in two directions. It does not have an inside and an outside in the way a perimeter does. The old security posture — keep the bad actors outside the wall — assumed a wall. The new posture has to assume a weave, and ask which threads it can afford to lose.
Three properties of the fabric are worth naming. First, it is differentiable: every interaction is a gradient signal that updates some model somewhere. Second, it is non-local: a signal collected in one jurisdiction is reasoned about in another. Third, it is recursive: the fabric reasons about its own reasoning, which means the audit trail you would need to govern it is itself a product of the thing being governed.
The question is no longer how to secure the conversation. It is how to negotiate with a fabric that has already absorbed it. Negotiation, in this context, is a technical discipline as much as a political one. It means designing interfaces that surface inference, not only output. It means giving the user a way to ask the fabric what it has concluded about them, and to contest those conclusions before they are acted upon.
Caribbean states have a particular interest in this conversation. We are small enough to be entirely modelled, and far enough from the centres of compute to have no meaningful say in how the modelling is done. The fabric will arrive here whether we participate in its weaving or not. The question is whether we arrive at the loom in time to leave a thread of our own.
— Fazal Ali · 16 May 2026 —
Next entry · Critical Theory in AI
Selecting agents and shielding your AI chats
No one can shield their AI chats from prosecutors pursuing charges. Conversations with AI assemblages could be demanded by adversaries with the patience of a state.
