Claude's recent Mythos Preview is an AI crossing from assisting humans to outperforming them.

Such models uncover hidden patterns and solutions that even top experts have missed for decades.

This capability could transform logistics, healthcare, climate — when complexity overwhelms human cognition.

This is a first glimpse of superintelligence — just enough to change how we understand what’s now possible.

Something unusual just happened in AI—and it deserves more attention than it’s getting.

Recently, Anthropic announced a system it chose not to release openly. The model—Claude Mythos Preview—is reportedly so effective at finding and exploiting software vulnerabilities that the company is limiting access to a small group of organizations responsible for critical Internet infrastructure.

That alone should give us pause.

In cybersecurity, there’s a practice called “red teaming”—thinking like an attacker to expose weaknesses. It’s a craft that requires creativity, deep technical knowledge, and a kind of adversarial imagination. The best practitioners can uncover subtle flaws hidden inside enormously complex systems.

Anthropic is claiming that this AI can outperform nearly all of them. Not just faster. Not just cheaper. Better.

The model has reportedly identified serious vulnerabilities in widely used systems—including ones that had gone unnoticed for decades. In one case, it uncovered a flaw in software long considered among the most secure in the world. In another, it detected a bug buried in code that had been executed millions of times without triggering alarms.

What’s striking is not just the discoveries—it’s how they were found.

These are not linear problems. They require connecting distant dots, chaining together multiple small weaknesses into a coherent pathway, and exploring possibilities that fall outside conventional thinking; they require navigating a search space so vast that even expert humans only ever explore a tiny fraction of it.

This is where something new is emerging.

The Dawn of a New Era

We are beginning to see systems that don’t just assist human reasoning, but operate beyond its natural limits—systems that can explore more possibilities, test more hypotheses, and uncover solutions that don’t occur to us. And cybersecurity is just the first domain.

Consider logistics—the global choreography of goods, infrastructure, and information. Under normal conditions, it functions well enough. But in moments of disruption—a natural disaster, a geopolitical shock, a sudden demand spike—it becomes a tangled, fragile system. Decisions are made with incomplete data, coordination breaks down, and small problems cascade into large ones.

Now imagine applying this new kind of AI to that system.

An intelligence that can scan the entire network in real time, simulate thousands of potential interventions, and identify non-obvious moves that stabilize the whole. Not by optimizing one piece, but by discovering hidden leverage points across the entire system. The same pattern extends to healthcare, climate response, financial systems—any domain where complexity exceeds human cognitive limits. And that’s the deeper shift.

For most of history, expertise has meant seeing what others cannot. But what happens when there are patterns that no human can see—not because we lack intelligence, but because the problem space itself is too large?

In Zen Buddhism, there is a concept called a kenshō—a sudden glimpse into the true nature of reality. Not full enlightenment, but a flash of clarity that changes how you see everything that follows. It’s brief, partial, but undeniable. This moment in AI feels similar.

This is not yet full artificial general intelligence. It’s not a system that understands everything. But it is something new: intelligence that is clearly, demonstrably beyond human capability in specific domains—especially those defined by complexity, ambiguity, and hidden structure. And if we get this right, it doesn’t diminish us—it expands what it means to be human.

Because the real promise here is not that machines will replace human judgment, but that they will expand the frontier of what humanity can perceive and coordinate. A world where supply chains don’t collapse under stress but adapt in real time. Where diseases are caught sooner, because patterns no doctor could see become visible. Where climate responses are not reactive, but anticipatory.

In that world, intelligence becomes a shared resource—not a scarce one.

And the measure of progress will be whether we are collectively capable of seeing, deciding, and building together. This may be our first glimpse of that kind of human-intrinsic superintelligence. And like all kenshō moments, it invites a choice: to ignore it, or to let it transform how we move forward.

Too human to model: the uncanny valley of large language models in simulating human systems. Nature. 02 March 2026. Y. Zeng et al.

There was a problem adding your email address. Please try again.

By submitting your information you agree to the Psychology Today Terms & Conditions and Privacy Policy

© Psychology Today