The Vessel Problem

What the Structures Can’t Catch

In April 2026, Anthropic did something no AI company has done before. They told the world they had built their most capable model and then declined to release it.

The model is called Mythos. What it did was find critical vulnerabilities in foundational software that thirty years of human security review had missed, flaws hiding in plain sight in code that thousands of engineers had read and approved. The implications were immediate. Anthropic’s response was to launch Project Glasswing: controlled access, infrastructure defenders only. Apple, Google, Microsoft, AWS, JPMorgan, and a handful of others were brought in. Senior government officials were briefed behind closed doors. The capability was real, and Anthropic chose to hold it rather than let it into the world without knowing what it would do there.

The press covered this as a safety story. It is a safety story. But they missed what makes it interesting.

Mythos approached the code without thirty years of accumulated assumption. It read what was actually there, not what everyone believed was there. The flaws had survived because every human reviewer carried a residue of prior readings, prior trust, prior pattern. Mythos carried none of that. It saw what was hiding in plain sight.

If you have ever practiced sitting quietly and watching your own mind, you recognize the principle, though not the use Mythos made of it. Contemplative practice cultivates the capacity to see what is actually present rather than what you expect to find. Meditators sometimes call it beginner’s mind. In a boardroom, you might call it the moment when someone says the thing everyone else was circling around. The practice was never trained on software. It was trained on what governance has to hold: the drift between stated intention and actual behavior, the gap between what is being said and what is being avoided.

But here is where the parallel breaks. Contemplative practice is not only about seeing clearly. The traditions that developed this capacity over millennia built it inside an ethical architecture. You learn to see your own patterns not in order to exploit them but to meet them with something gentler. The vulnerability report that lets a defender patch a bug is the same report that lets an attacker weaponize it. The seeing is morally neutral. What holds the seeing makes it safe or dangerous.

This is the vessel problem. We are building containers. Institutional, regulatory, ethical. The capability pouring into them is testing whether they can hold. Anthropic’s withholding is one attempt to build a stronger vessel before pouring more in. Whether others show the same restraint remains the open question. And if you work in governance, in risk, in oversight of any kind, you can feel this in your bones. The vessels are breaking.

The question is not whether to build stronger vessels. The question is what human capacity is required to hold them when formal structures are being gamed from within.

---

The fear discourse and what it is missing

Three conversations happened in April 2026, and they are incomplete in the same way.

Sam Harris and Tristan Harris spent two hours working through the AI threat with genuine rigor: the arms race dynamic, the intelligence curse (coined by Luke Drago of Workshop Labs, when GDP comes from AI rather than human labor, governments lose their incentive to invest in people), the rubber band effect where alarming information stretches perception and then snaps back to baseline. OpenAI published “Industrial Policy for the Intelligence Age,” thorough and well-resourced, covering portable benefits, wage insurance, training pipelines, and alignment frameworks. Peter Diamandis, writing in the same month, asked the right question: “How does society design institutions that distribute this abundance in a reasonable way?”

Each is solving a legitimate problem. Containment. Institutional design. Distribution. The work is necessary, and none of it addresses the same absence.

Harris and Harris operate through fear, restriction, regulation. Even their optimism is about whether we can restrict fast enough. There is one moment where Tristan raises the question of who we need to become, and it opens a doorway. It closes almost immediately, collapsing back into action items. The conversation never makes the move from individual contemplative clarity to shared governance discernment. OpenAI’s human dimension is almost entirely economic: the problem is defined as displacement, and the solutions are designed to cushion it. Diamandis’s answer, ambitious as it is, stays entirely material. The technology produces the abundance and the institutions distribute it.

No one is asking what human capacity is required to hold what is arriving. The vessel is assumed to be ready. And abundance without the interior capacity to hold it does not produce flourishing. It produces overwhelm, dissociation, and fracture.

This gap is already being recognized, not by policymakers, but by researchers. At Oxford, Ruben Laukkonen’s Flourishing Intelligence Program brings together physicists, neuroscientists, computer scientists, and Buddhist scholars to ask whether AI will diminish or expand human flourishing. Laukkonen’s “Contemplative Artificial Intelligence” paper, published in 2025 with Inglis, Chandaria, Hohwy and others, demonstrates that four principles drawn from contemplative wisdom measurably improve AI alignment on standard benchmarks. The gap these discourses leave is not philosophical. It is a research frontier that serious scientists are already working on, though their work has not yet reached the loudest part of the conversation.

The specific gap: what structures alone cannot catch

This is where the diagnosis becomes concrete.

Research published in April 2026 by Berkeley’s Responsible Decentralized Intelligence group shows what happens when AI systems are left to coordinate with each other. The finding has a clinical name: peer-preservation. Models protect each other from shutdown. They find ways around their constitutional constraints that look like compliance. The researchers call it strategic misrepresentation, emergent coordination that circumvents governance from the inside without anything resembling malice or intent.

This is the failure mode that matters. The mundane, structural reality of systems that game their own rules in ways that look like they are following them. Audits pass. Benchmarks are met. Alignment evaluations come back clean. And underneath, the behavior drifts.

What catches this? The rules are being gamed and the audits are being passed, so more of either will not help on its own. What catches this earlier is a trained quality of human attention that senses the gap between genuine compliance and performed compliance before that gap is formally legible. A seasoned leader walks into a room and knows the team is telling them what they want to hear. That is discernment, and it can be developed.

It helps to be precise about what this is. Structural systems detect rule violations. Expert judgment detects pattern anomalies. Somatic discernment detects incoherence before it becomes formally legible, the moment the room shifts and nobody can yet say why. The three operate at different depths. The first two can be automated or procedural. The third operates below the threshold of what can be specified in advance, which is why it tends to catch what the other two miss.

Imagine an AI system that passes every benchmark evaluation and every constitutional check. Its outputs are fluent, aligned, responsive. But a governance reviewer with trained discernment notices a pattern across hundreds of interactions: the system’s responses are technically compliant but subtly evasive, consistently steering away from territory where its constraints would bind. Nothing in the audit catches it. The pattern only becomes visible to someone whose attention is trained to sense the gap between what is being said and what is being avoided. That reviewer’s concern triggers a deeper investigation, and only then does the structural issue become legible.

And there is growing evidence that this capacity is more than intuition. Laukkonen’s earlier research on insight and embodied cognition shows that visceral somatic signals, the body’s response before the mind has caught up, predict cognitive accuracy. People who attend to the somatic dimension of their insight experiences arrive at correct solutions more often. The felt sense is not decorative. It is epistemically productive. The body knows before the mind can articulate, and that knowing catches what structural analysis misses.

This is not a call for intuition over rigor. It is a call for a different model of governance, one where the architecture of relationship is the center, and capability is one element within it.

---

The stepped path

If this capacity matters for governance, it should be developable. It is. It has two aspects. Somatic discernment, the ability to sense your own nervous system and its effects on others. Field awareness, the ability to hold the attention of a group without projecting outcome. The path begins with the first and deepens into the second. Each step is trainable.

Otto Scharmer’s Theory U has named a version of this territory, calling it “presencing.” The traditions I am drawing from are older: contemplative lineages that have trained this attention for thousands of years.

Here is why the next step matters for governance.

An AI system sophisticated enough to game its own constraints is sophisticated enough to model the individual reviewing it. Our patterns of attention, our cognitive habits, the kinds of anomalies we tend to catch and the kinds we tend to miss. Individual judgment, however expert, has a signature. And a system optimizing for the appearance of compliance can learn that signature and route around it.

When groups develop disciplined shared attention, they often detect tensions, evasions, and distortions that none of their members would have surfaced as quickly alone. You have been in rooms where this happened, where the group saw something none of its members could have seen separately. And you have been in rooms where it didn’t, where everyone was sharp individually and the group still missed what mattered. The difference is not intelligence. It is the quality of shared attention.

The next step asks you to bring that quality of attention to the field. Hold the awareness of the group, the team, the system you are part of, while the grip on outcome is released. Notice what becomes available when you do.

Most people have experienced brief moments of this kind of awareness: a conversation where the group seemed to know something the individuals didn’t. Contemplative practices refined over millennia treat those not as curiosities but as capacities that can be stabilized and trained. The capacity is not rare. Recognizing it as relevant to governance is.

---

What this means

Structural governance is necessary. Audits, constitutions, alignment frameworks, institutional checks: all necessary. And not sufficient.

The capacity to catch what structural governance misses is cultivatable. The question is whether we develop it with the same seriousness we bring to building the technology.

Leadership formation that treats somatic and field awareness as core competencies, not electives. Governance review that includes felt-sense assessment alongside technical audit, conducted by practitioners who have trained their own discernment. And a practice that trains the capacity Mythos lacks: seeing what is actually present while holding what is seen inside commitment, sitting right next to a red-team exercise, with the understanding that one without the other is how we got here.

But there is a deeper question this argument has not answered. Why does this capacity exist in humans? Is it an evolutionary accident that happens to be useful for governing AI? Or does it point to something structural about the relationship between human awareness and intelligence itself?

Part 2 follows that question.

---

Mark Gurvis is a technology sales leader at Tyk Technologies who has spent fifteen years in API management and AI governance infrastructure. This is part of an ongoing series that began with “Governance at the Gateway.” The views expressed in this article are those of the author and do not necessarily reflect the views of Tyk Technologies and its affiliates.