The Monster in the Mirror: Anthropic's Glasswing and the New Architecture of Global Power

1A Capability Unlike Anything Before

First, understand what Mythos actually is. This is not incremental progress.

In an internal document accidentally or purposely made public, Anthropic called the model "by far the most powerful AI model we've ever developed." That framing is worth taking seriously, even discounting for promotional language. Mythos can chain together four or five vulnerabilities that appear harmless independently into a single sophisticated exploit. It wrote a browser exploit using a complex technique that escaped both renderer and OS sandboxes. It generated working exploits with little or no human steering. Anthropic's own disclosures show it outperformed its next best public model by a wide margin on internal cyber benchmarks.

Anthropic itself has warned that Mythos "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."

They built it. They know exactly what it can do. And they have decided, alone and without any external process, who gets access to it, on what terms, and at what price. That is not just a product decision. It is the beginning of a governance decision. Right now, that governance is being made entirely by the capability holder itself.

2The Governance Vacuum Nobody Is Filling

We once had something resembling external governance in technology. GDPR, imperfect as it was, imposed the same rules on every company regardless of size, origin, or intent. Critically, the rules were not written by the companies being governed.

What we have in frontier AI today is the complete inverse. The most powerful actors write their own rules, enforce their own rules, and adjudicate their own compliance. Anthropic's responsible scaling policy, safety frameworks, and Glasswing partner criteria are all self-authored. There is no external body, no international agreement, and no GDPR equivalent for frontier AI capability.

Anthropic has already shown what self-governance can look like in practice. When OpenClaw, a popular open-source AI agent tool, became inconvenient, Anthropic blocked it from Claude subscriptions. Whatever the merits of that specific case, the broader lesson is clear: when a company writes its own rules, it enforces them in its own interest. There is no appeals process and no independent arbiter.

3Who Decided, and How

The Glasswing partner list includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Beyond these, roughly forty additional organizations have been granted access.

Anthropic has not published a transparent, externally auditable framework explaining how organizations qualify, how geographic balance is handled, what public-interest obligations partners assume, or what recourse exists for those outside the initial circle. There was no open process and no methodology anyone outside Anthropic could scrutinize.

The Linux Foundation's presence is genuine and worth noting. But the rest of the list is dominated by American companies with major commercial interests in the same infrastructure Mythos is scanning. And the geography may be the sharpest issue of all: Asia, Africa, Latin America, and much of the Global South are absent from the table despite running the same operating systems, browsers, and open-source infrastructure.

4The Two-Tier World Being Built

Anthropic has said it will not make Mythos generally available. What it plans to release instead are advanced versions of its Opus model line, capable but deliberately less powerful in the capabilities that matter most. Mythos finds and fixes vulnerabilities in global infrastructure. A less capable version gets released to everyone else.

US companies get to fix their house. The rest of the world does not yet know which rooms in their house are on fire.

There is a second, less visible dimension to this asymmetry. Anthropic says over 99 percent of the vulnerabilities Mythos has found remain undisclosed because patches are not yet ready. That approach is understandable, but it also means the actor with the strongest discovery capability controls the timing, sequencing, and information asymmetry around global patching.

5The Business Model Nobody Is Talking About

Training a frontier model at Mythos's scale costs hundreds of millions of dollars. Anthropic has publicly stated it will not make this model generally available. So the obvious question is: where is the return on that investment?

• Glasswing partners may be paying for premium access at enterprise scale.
• Mythos may function as Anthropic's internal engine, accelerating future model development and preserving a durable lead.
• The gap between Mythos and publicly available models may itself be the product: a permanent capability differential that benefits select partners while everyone else receives a weaker tier.

We do not have direct proof of intent. But we do have the economics, the access structure, and a public statement that a model of this scale and cost will not be generally released. Those facts deserve a direct explanation that has not yet been given.

6Whose Safety, Exactly?

Dario Amodei deserves real credit for refusing Pentagon demands in early 2026 to allow Claude to be used for mass domestic surveillance of American citizens or fully autonomous weapons. That took conviction.

But Amodei has also written publicly about an "entente" strategy, a coalition of democracies using AI to achieve decisive military and strategic superiority over adversaries while distributing benefits to cooperating nations. He has stated that democracies have a legitimate interest in AI-powered military and geopolitical tools, and that lawful foreign intelligence and counterintelligence uses are acceptable.

So when Mythos is simultaneously finding thousands of zero-days in global infrastructure and Anthropic is in active dialogue with US government and defense officials, the question is not paranoid. It is logical. Who benefits from discovering these vulnerabilities, and who decides whether they get disclosed, patched, or quietly retained?

7The Drift Toward Quasi-Sovereignty

Companies like Anthropic do not merely build tools anymore. They decide access, define acceptable use, coordinate with states, restrict adversaries, shape industry norms, and manage sensitive knowledge about vulnerabilities affecting the wider public. Those functions begin to resemble pieces of governance.

Yet unlike states, such firms are not elected. Unlike regulators, they are not structurally accountable to the people affected by their decisions. Unlike treaty institutions, they do not represent a negotiated international consensus.

Even if Anthropic's founders remain sincere in their stated values, they cannot hold that line indefinitely against the gravity of state power. The company is incorporated in the United States, subject to US law, dependent on Amazon's cloud infrastructure, and reliant on capital at a scale very few sources can provide. Capital is leverage. Infrastructure is leverage. Law is leverage.

The Real Question

I am not arguing that Anthropic is acting in bad faith. I am not arguing that finding and disclosing vulnerabilities is wrong. I am not arguing that some degree of caution around a model this powerful is unreasonable.

I am arguing that a private company, with no democratic mandate and no external accountability, has built one of the most powerful cyber capabilities in history. It is deciding unilaterally who accesses it, which tier of capability the rest of the world deserves, and how it interfaces with state power. The rest of the world, which runs on the same infrastructure, faces the same vulnerabilities, and will live with the consequences, was not consulted.

Without that, we are not just building safer software. We are quietly accepting a new architecture in which a private firm in San Francisco decides how much security the rest of the world gets to see, when it gets to see it, and on whose terms.