Mythos: The AI Anthropic Built and Won't Let You Use

Anthropic just announced the most capable AI model they have ever built. You cannot use it. Neither can most companies. Neither can the Pentagon — and that one is not an accident.

The model is called Mythos, and the only reason any of us know about it is because the documentation describing it leaked from one of Anthropic's own unsecured servers in late March. The story since then has been one of the strangest product launches in AI history: a frontier model whose existence was confirmed only after a data leak, whose rollout has been deliberately starved of customers, and whose announcement crashed cybersecurity stocks in a single trading session.

If you want to understand where frontier AI is actually heading in 2026, this is the story to pay attention to.

What we actually know

Here are the confirmed facts, drawn from CNBC, Fortune, and the Times of India coverage over the past two weeks:

• Anthropic has been internally testing a new model called Mythos with a small group of early-access customers since at least mid-March.
• Existence of the model was disclosed only after a data leak from an unsecured Anthropic data store revealed internal documentation.
• Anthropic has explicitly described Mythos as a "step change" in capability over previous Claude releases — that's the company's own framing, not analysts'.
• Public release is being held back. Instead, Anthropic is rolling Mythos out via a closed program called Project Glasswing, restricted to a hand-picked set of cybersecurity firms.
• The day the news broke, the iShares Cybersecurity ETF (IHAK) and several major cyber stocks dropped as investors digested the implication: an AI capable of running offensive cyber operations at this level changes the threat model for every company in the sector.
• All of this came weeks after Anthropic's public clash with the Department of Defense, which had previously chosen Anthropic for AI work and then publicly walked back the partnership over safety disagreements.

That is a lot to absorb. Let me break the parts that actually matter.

"Step change" is not marketing

When a frontier AI lab calls something a "step change," they almost never mean it the way a marketing team would. Step change is an internal evaluation term — it usually means a model crossed a threshold on a benchmark or capability that the prior generation could not touch at all. Not 10% better. Not 30% better. A new category.

Past examples of "step change" framing inside frontier labs have lined up with things like multi-step autonomous reasoning, novel scientific synthesis, or agentic tool use that worked end-to-end without supervision. Whatever Mythos is doing on the cyber side, Anthropic believes it crossed a line that previous Claude models — including Opus 4.6 — could not.

That is the core reason this story is not normal tech news.

Project Glasswing is the real story

Most people are reading the headlines as "Anthropic built a scary model and won't release it." That misses what's actually happening.

Project Glasswing is not a freeze. It is a controlled deployment. Mythos is being given to customers — just very carefully chosen ones, almost all of them on the defensive side of cybersecurity. The bet Anthropic is making is that the only sane way to deploy a model this capable in the cyber domain is to put it in the hands of defenders before attackers can build their own version.

This is a meaningful shift in how a frontier lab is treating release strategy. The previous default was: ship to API customers, gate sensitive capabilities behind policies, and hope misuse stays manageable. Glasswing is closer to a vaccine rollout — pick the highest-leverage population first, build up immunity in the system, then widen access.

Whether you think that's responsible or paternalistic depends on your priors. But it is genuinely new.

Why the cyber stocks dropped

Cybersecurity ETFs do not move on AI announcements. They move on threat-landscape news. The fact that they dropped on the Mythos report tells you what professional analysts think Mythos can do.

If a model can plan and execute cyber operations end-to-end at superhuman speed, it does not just make attackers stronger. It makes the entire labor model of defensive cybersecurity less defensible. A SOC analyst is suddenly competing against an attacker who has infinite patience, infinite parallelism, and zero human cost. Even if the model is locked behind Glasswing, the existence of capability at that level reprices the whole industry's risk.

That is why the market reaction was specific and immediate, even though no one outside Anthropic has actually used the model yet.

The Pentagon backstory matters

Easy to forget in the news cycle: just weeks before Mythos surfaced, Anthropic and the Department of Defense had a very public falling out. Anthropic had been the Pentagon's preferred AI partner. Then a court ruling and a safety dispute pushed the relationship into open conflict, and the partnership was effectively suspended.

Now consider the timing. Anthropic — the lab that just walked away from the largest defense customer in the world over safety disagreements — is the same lab that has now decided a model is so capable they will not release it through normal channels.

If you are an Anthropic skeptic, you read this as a company trying to relaunch its safety brand after a bruising fight. If you trust the lab, you read it as exactly what they have always said they would do: catch a dangerous capability before deployment and deliberately constrain rollout. Both readings are coherent. Neither is provable from the outside.

The real question

Almost every take I have seen on Mythos so far has fixated on the wrong question. People are asking "is Mythos dangerous?" That question is almost meaningless. Of course it is dangerous — a frontier model with novel cyber capability is dangerous by definition, the same way a new exploit is dangerous before it is patched.

The question that actually matters is who gets it first. Glasswing decides that. The handful of companies inside that program will have access to a class of AI capability that no one else does. Their products, their detection systems, and their threat intelligence will pull ahead of everyone outside the program in a way that compounds quickly.

If you work in cybersecurity, on either side, this is the moment to watch. If you work in AI policy, this is the moment to ask hard questions about who gets to be on the "vaccine rollout" list and how the criteria are set. And if you are building anything that depends on cyber risk staying static — your threat model just got more interesting.

What I am watching next

Three things will tell us whether the Mythos story is the start of a real shift or a one-off:

1. Does Glasswing expand? A controlled rollout that never widens is just a moat. A controlled rollout that scales is a new release model.
2. Do other labs follow? If OpenAI and Google adopt similar staged-deployment programs for their next frontier models, Anthropic just set a new industry default.
3. Does anything leak from inside Glasswing? Capabilities this strong, distributed to multiple companies, are very hard to keep contained for long.

For now, the most honest summary is this: Anthropic built something they themselves are not ready to ship into the open, they got caught building it before they were ready to talk about it, and they are now improvising a release strategy in public.

That is a much more interesting story than "scary AI." And it is the story I think we will be unpacking for the rest of the year.

---

Sources: CNBC — Anthropic limits rollout of Mythos AI model, Fortune — Anthropic 'Mythos' AI model representing 'step change', Times of India coverage of Project Glasswing, March–April 2026.