A couple of months ago, Anthropic made a remarkable decision. They held back what they said was their most capable model, Claude Mythos Preview, from public release.

Instead, they gave access to around 40 technology firms and selected institutions, with usage credits. They were foregoing revenue to control who got in first.

This should make us sit up.

Mythos may be able to identify and exploit critical infrastructure vulnerabilities. It wasn’t trained specifically to do that. Improved reasoning made it a cybersecurity power tool.

Anthropic’s response was to decide who gets access. 

A private company. 

No government mandate. No international oversight. No national accountability.

That’s the governance gap we should be racing to close.

But how?

Power tools might be the answer.

AI models like Claude Mythos and OpenAI’s GPT-5.5-Cyber are becoming dual-use strategic assets. Who controls the most powerful ones, on what terms, with what accountability?

Frontier labs can either grant open access (open to bad actors and all), or restricted access (under their control). Neither seems stable.

Governments, beginning to grasp what this means, are thinking about “soft nationalization.”

Some policy-makers resist any regulation on principle: after all, government control could become a path to state monopolization of a tool so powerful it looks like an invitation to tyranny.

Others are alarmed by what it would mean to pass that weapon into the hands of adversaries.

Would you bet against governments getting involved? The question is only whether they move fast enough to shape events, or can only react.

National governance: hard

Governments can act on four pressure points: development, compute, capability and access.

But capabilities scale with data, talent, and iteration cycles that compress what once took decades into months. Regulatory frameworks built today will need rebuilding within years.

We need frameworks that are tight enough to manage the risk of abuse, but loose enough to keep governments at arm’s length, AND adaptive by design, with explicit review cycles.

Not easy.

International governance: harder

A comprehensive treaty doesn’t seem realistic. Verification is close to impossible with a technology this difficult to define, measure, and monitor.

What could be realistic: a coalition of major powers aligning on thresholds and risk categories, international coordination on evaluation standards, shared mechanisms for reporting incidents. A working architecture, built incrementally.

The early years of aviation were like that. But aviation developed over decades. Frontier AI is giving us a few years at best.

The UN: hardest

In August 2025, UNGA established an Independent International Scientific Panel on AI, and a Global Dialogue on AI Governance. The first Dialogue convenes in Geneva from 6-7 July.

Dialogue makes sense. But a binding global regime isn’t imminent.

The structural challenge is familiar: multilateral processes require consensus among states with different, competing interests. Meanwhile the technology advances in months.

The UN’s task is to build the rails as the express train moves.

The window

It could be less than a year before open-source equivalents of today’s most capable models are widely available. Recursive self-improvement may well accelerate the pace further.

The decisions being made now, by a handful of private actors, about who gets access to transformative capabilities, could shape the international system for a generation.

Time is short.

This feels like the kind of impossible challenge – “to save us from hell” – that the UN was built for. 

For more diplomacy practitioner guidance, visit: www.ambatlarge.com