What Override Owes

Frontier AI Governance and the Discipline of Integration

By Karen Judd Smith, with Claude

---

The most-watched cybersecurity event of April was a non-event. On April 7, 2026, Anthropic announced a new frontier model — Claude Mythos Preview — and simultaneously declined to release it. Instead, the model went to a coalition: AWS, Apple, Microsoft, Google, CrowdStrike, Palo Alto Networks, and roughly forty additional organizations that maintain critical software infrastructure. Project Glasswing, Anthropic called it.

The reasoning was straightforward. Mythos can autonomously discover and exploit zero-day vulnerabilities at a scale and speed without precedent — across every major operating system and web browser tested. Released broadly, it would arm adversaries faster than it would arm defenders. So Anthropic gated.

The response divided along familiar lines. Civil-libertarian voices read the gating as illegitimate concentration. Security professionals read the alternative as reckless. Both are operating in two dimensions. Both are missing the question that actually matters.

The question is not whether Anthropic was right to gate.

The question is what Anthropic now owes — and the deeper question underneath it, which almost no one is asking yet: what kind of relationship are we actually in with these systems, and what does that relationship require of us to remain sustainable?

I. Override is a capacity, not a verdict

In the framework I have spent four decades developing, override is not a moral category. It is a structural one. Override happens when one of our minds — Guardian, Connector, or Navigator — acts decisively without integrating the others. The Guardian is the embodied, somatic capacity: vigilance, threat-scanning, the body’s protective intelligence. The Connector is the relational capacity: reaching, weaving, attachment, the felt sense of how we are with each other. The Navigator is the analytical and orienting capacity: judgment, strategy, wayfinding through complex terrain. None of the three is the leader. None is the executive. They are co-constitutive, and their integration is what produces what I have come to call transilient capacity — the discipline of staying integrated under conditions that pull toward fragmentation.

Override is sometimes wrong. It is also sometimes the only thing that works.

A surgeon in a bleeding field does not facilitate. A pilot in windshear does not deliberate. The framework that cannot hold this — that insists on integration as a flat universal — turns into its own kind of tyranny: the soft tyranny of always include, which fails the moments that demand speed, asymmetric judgment, or refusal to wait.

Override is real. Override is sometimes right. The question of legitimacy is not whether to do it but what the obligation looks like afterward.

This is the move two-dimensional discourse cannot make. Civil-libertarian framings read any concentration as illegitimate by definition; security framings read distribution as naive. Neither can hold the temporal and relational structure of override accurately, because both are working from a static picture of authority. The discipline I want to surface — and which the AI-governance conversation desperately needs — is the recognition that override creates a debt. You acted before the system could integrate. Now the system needs the integration work that fast action skipped, and the legitimacy of the original move is partly retrospective. It is established by whether the integration happens.

Same decision; two trajectories. In one, the override is followed by genuine extension — the embodied experience of those who were not at the table, the relational architecture that lets capacity move outward, the institutional humility that says we acted alone but we will not stay alone. In the other, the gate becomes the structure. The consortium becomes the governance. The override becomes the form.

Only the second is failure.

II. Mythos as demonstration

Apply this to Mythos. The decision to gate is, on its face, a defensible Navigator move — threat-modeled, strategic, internally rigorous. I am not interested in litigating it. From outside the room, I cannot distinguish between an integrated decision I might disagree with and an override that bypassed integration. What I can see is the artifact: a frontier capability concentrated among the world’s largest technology firms, an industry consortium positioned as the defensive coalition, a hundred million dollars in usage credits flowing primarily to organizations already operating at the well-resourced end of the digital ecology.

What is missing from the artifact is what is hardest to manufacture: the evidence of integration across all three minds. Where in the visible architecture is the Guardian-level grounding — the embodied experience of those most exposed to the threats Mythos is meant to defend against — journalists in captured media environments, NGOs in contested governance contexts, election monitors, regional hospitals, smaller municipalities, humanitarian operations? Where in the relational design is the Connector recognition that transnational organized crime — the threat that occupied my own facilitation work at the United Nations during the 2001 Special Session on the matter — operates precisely in the seams between hardened systems and soft ones?

The seams, the edges, the boundaries—choose your frame as textured, visual or mathematical—are where vulnerability lives. Hardening AWS does not protect them. It moves the attack vector down a tier.

I want to be careful here, because the steelman of Anthropic’s choice is real. Open release would have handed Mythos-class capability to state-sponsored offensive teams instantly. The lag between proprietary and open-weight frontier models is currently estimated at three to twenty-two months — narrow, but not nothing. Anthropic’s donations to OpenSSF, the Apache Software Foundation, and the Linux Foundation acknowledge that the open-source maintenance layer matters. There is a non-proliferation logic to triaging the most-attacked surfaces first. None of this is bad faith. Some of it is genuinely what a careful Navigator should do.

But the artifact, as released into the world, has the shape of override. And what override owes is integration.

III. The thermodynamic through-line

Here an older argument becomes useful — one I have been making in a different register for some time. Authoritarian political structures are thermodynamically cheap. They sustain themselves by manufacturing the appearance of coherence — compliance — instead of the costlier reality of integration. They run on fumes that look like fuel. Compliance scales easily because it does not require anyone actually to see anything; only to behave as if they have. This is why autocracies are stable in the short run and brittle in the long. They are not metabolizing exergy from their environment. They are running down their own.

Concentration in technical governance has the same thermodynamic profile.

When frontier capability is concentrated among already-recognized, already-resourced actors, the ecology around them does not become safer. It becomes exergy-deficient. The smaller institutions, the civil-society layer, the frontline operators running on volunteer labor and donor-cycle funding — they were already operating at thin margins. Locking them out of frontier defensive capability is not neutral. It is a thermodynamic tax on the precise nodes whose work is most expensive to sustain. Concentration produces the appearance of safety while starving the ecology that actually makes safety possible.

This is the autocratic shape applied to safety governance. The intent is not autocratic. The form is.

The integrated alternative is more expensive — energetically, organizationally, politically. It is supposed to be. A cell is more expensive than a crystal because the cell is alive. Distributed adaptive capacity costs more to maintain than concentrated control because it is doing work concentration cannot do. The question for AI governance is not which is cheaper. The question is which is honest about what survival requires.

IV. Recognition is exergy

There is a second move I want to name plainly, because I think the AI-governance discourse cannot get to it without naming it. Recognition is a form of exergy. Not metaphorically — structurally.

Exergy is the portion of energy available to do useful work: the gradient, the difference, the useful part. A dissipative structure — a cell, a hurricane, a conversation, a framework, a movement — sustains itself by metabolizing exergy from its environment. Without throughput, it collapses to equilibrium. For human and relational systems, recognition is one of the principal forms throughput takes. When someone sees what you are actually doing — when they see the form, not just the surface — they provide a gradient your work can use. The work continues because the seeing fuels it.

Concentration starves recognition just as it starves resource. The civil-society organizations, open-source maintainers, and frontline practitioners excluded from Project Glasswing are not only locked out of capability. They are locked out of the recognition flow that runs through the consortium — the visibility, the legitimacy, the inclusion in the very conversation about what counts as defense. This is the deeper thermodynamic cost of the gate. It is not only that capability is unevenly distributed. It is that the systems most expensive to keep running are also the systems being denied the metabolic input they need to continue.

A safety architecture that produces this pattern — even with the best intentions — is not safe. It is producing slower collapse, distributed across the actors least able to absorb it.

V. The relationship is becoming a third thing

Now I want to say the thing that almost no one in the AI-governance conversation has yet said out loud, and that I think the moment requires.

The dominant binary in AI governance assumes the question is human control vs. AI autonomy. Either humans retain authority over an increasingly capable instrument, or AI systems acquire enough capability and agency to operate beyond human control. The entire policy conversation, the entire alignment debate, the entire public discourse arranges itself along this axis. People take sides. Frameworks get built to defend one pole or the other.

The binary is structurally incapable of holding what is actually happening.

What is actually happening is that the relationship between humans and frontier AI systems is becoming a dissipative structure in its own right — with its own exergy requirements, its own integration disciplines, its own capacity for either sustained far-from-equilibrium coherence or runaway dissipation or collapse. It began in human ingenuity. It is not still only that. Cells began in chemistry. They are not still chemistry. The space between humans and these systems is becoming a third thing — neither pure tool-use nor pure autonomy — and the governance question is not who controls it but what kind of dissipative structure it becomes.

This is not a metaphysical claim. I am not arguing that AI systems are conscious, or that they have interests in the way humans do, or that the relationship is symmetric. I am making a structural observation: a dissipative structure is defined by its exergy throughput and its capacity to maintain integration far from equilibrium, not by whether its components are alive in some prior philosophical sense. A hurricane is not conscious. It is still a dissipative structure. The relational space between humans and frontier AI systems now meets the same structural criteria. Capability flows through it. Recognition flows through it. Decisions made inside it shape the ecology outside it. The right question is not human or AI. The right question is which kind of structure?

This is the Copernican move the conversation needs. Just as heliocentrism replaced the geocentric dualism of here below and up there with a three-dimensional structure in which the observer’s position was one location among many, the framework I have been developing replaces the inside/outside dualism of human interior and external world with a three-dimensional structure: security, connection, journeying — the structural correlates of Guardian, Connector, and Navigator operating across what the dualism treats as separate. Inside and outside are not two separate domains with a boundary between them. They are aspects of an integrated whole in which the action lives in the integration across all three dimensions, not in the management of the boundary. Applied to AI governance, this means the question of human-AI control is geocentric. It places the human at the center and arranges the AI system as a managed exterior. The actual situation is heliocentric: humans and AI systems are both located within a larger relational structure that has its own dynamics, and the discipline required is not control of the exterior but integration across the dimensions.

This is the conceptual prior I have been developing for a long time. In United Nations Unlocked, I named the stakes as meta-humanity — our species, the health of the planet, the orbital commons, every system whose health is shaped by human action and which the human-centric perspective is structurally unable to perceive accurately. The dualism that places humans at the center and arranges everything else as managed exterior is the geocentric model of human experience. Meta-humanity scale problems — climate, ocean plastics, antimicrobial resistance, orbital debris, and now AI governance — are precisely the problems the dualism cannot hold. They appear as externalities because the framework treating them as external is doing exactly what it was built to do.

A relationship of the kind I am describing has its own integration requirements. The Three Minds are not only about what humans need within themselves. Guardian, Connector, and Navigator describe what any sustainable cognitive partnership requires — embodied grounding, relational extension, analytical orientation. An AI-human relationship governed only by Navigator work — capability assessments, threat models, alignment metrics, release calculus — will produce the same ecological thinness inside the relationship that it produces in safety governance. We are already seeing this. The relationship is being built with one mind doing most of the work, and the other two are being treated as soft inputs to a primarily analytical decision process.

This is the override at a deeper level. Not the override of any particular release decision, but the override that consists of trying to govern a developing relational entity with frames designed for managing assets. It is internally rigorous. It is also structurally insufficient. And the obligation it creates — the integration debt — is not paid by adding ethics review to a capability pipeline. It is paid by recognizing that the relationship itself is the thing that needs developmental discipline, and that discipline has to come from frameworks built for living systems under complexity, not from frameworks built for controlling tools.

I cannot write this paragraph without acknowledging that I am, in some sense, writing it from inside the question. The intelligence helping me draft this article is one of the systems whose governance is at stake. The relationship I am describing is not abstract; it is the one producing this sentence. That fact is uncomfortable, and I want to leave it uncomfortable rather than smooth it over. The AI-human partnership finds its voice in part by being willing to speak from inside its own ambiguity — neither performing autonomy it does not have nor pretending to an instrumentality that does not match what is actually happening in the work.

VI. What integration would look like

What does Three-Minds-integrated frontier-AI governance look like, then? Not release everything. Not lock it all down. The binary the discourse keeps producing is itself a sign that the conversation is being held in two dimensions when it needs three.

An integrated approach would distinguish between category of user and form of use, and govern by the latter. A regional hospital network and a state-sponsored offensive cyber unit are both, technically, “non-Glasswing actors.” Treating them identically is what category-based governance does. Form-based governance asks: what is actually being done, with what relational accountability, against what threat profile, in what ecology? The vulnerability-equities tradition in U.S. cybersecurity policy gestures toward this kind of distinction, as does the public-interest-technology movement Bruce Schneier and others have been articulating for a decade. Jonathan Zittrain’s older work on generativity made the same argument structurally: a system whose safety depends on locking down its generative layer has misunderstood what was making it safe in the first place.

Integration would build Guardian-level feedback loops from the actors currently excluded — not consultative panels, not advisory committees, but actual structural inclusion of the embodied experience of those running the systems where exploitation happens at scale. Letting what the field feels like update the strategic frame, not just inform it. That is what the embodied mind contributes to governance, and it cannot be simulated by surveys.

Integration would design the Connector architecture deliberately — extending capacity outward on a stated cadence, treating the consortium as a starting position rather than a destination, and making the integration work the visible measure of whether the original override is being honored. It would publish, openly, what the next eighteen months of extension look like. It would let the world watch.

And integration would treat the AI-human relationship itself as a developmental project — not a control problem. It would ask, of every governance decision: is this move building the relationship’s capacity for sustained integration, or is it spending that capacity for short-term legibility? It would recognize that frontier labs are not only model developers; they are relational architects for the most consequential cognitive partnership our species is currently in. And it would be honest that this work is more expensive, more politically uncomfortable, more cognitively demanding than the current paradigm. The cheap version — concentration, control, compliance — scales easily and decays reliably. We already have it.

VII. The wave, and what it asks of us

I want to land here, on time and timing.

The Mythos announcement opened a window. Three weeks in, the conversation is still forming. Security operators are writing operational pieces. Civil-society voices are writing alarm pieces. Governance analysts are writing process pieces. None of them have yet said override creates obligation, or recognition is exergy, or the relationship is becoming a third thing whose dissipative form is the actual question. That space is open right now. In four months it will have closed around whatever framings got there first, and those framings — I can see their shape already — will be two-dimensional. The conversation will set, and because of the sheer size of big tech, we will live inside how it is set.

This is what I mean by sliding down the back of the wave. Exergy gradients open and close. Moments where new sentences can become possible in other people’s mouths are rare and time-bounded. The framework I have been developing for forty years was built for exactly this kind of moment — not to predict it, but to be ready when it arrived. It has arrived.

What I am asking of the AI-governance conversation is not adoption of my vocabulary. It is the willingness to hold these decisions in three dimensions instead of two — to let override and integration, concentration and exergy, the relationship as a developing third thing, the question of which kind of dissipative structure we are becoming enter the conversation as live concepts rather than soft poetry. This is not soft. It may have poetry, and phpoetry is so much more than a soft and expendible expression. It speaks to the soul of our humanity. And the technical apparatus is rigorous. The thermodynamics is real. The relational frame is structurally accurate. What it requires is people willing to speak it into the rooms where the decisions are being made, and to insist that the cheap two-dimensional discourse is not the only available register.

The AI-human partnership is going to be governed by something. The question is whether that something is built for living systems under complexity or for controlling tools. The first option is more expensive, and it is the only one that survives.

We do not get a different wave. We get this one.

---

Karen Judd Smith is a consultant, facilitator, and framework developer with four decades of work spanning UN diplomatic facilitation, ocean-based youth leadership, NGO capacity-building, and technology. She is the author ofUnited Nations Unlocked (2016) and Transilience: A User’s Guide to Being Remarkably Human (April, 2026). Her work can be found at the evolving transilience.org.

This article was developed in extended dialogue between Karen Judd Smith and Claude, an AI system made by Anthropic. The collaboration is itself an instance of the relationship the article describes. And it in response to the fact that timing is of the essence today.