March 28, 2026
What the Mythos leak revealed before the launch
Before Anthropic announced Project Glasswing, an exposed draft showed how the company was thinking about capability, cybersecurity and access.
Update, April 2026: Anthropic subsequently announced Mythos Preview and Project Glasswing. This article is preserved as the analysis I wrote from the exposed material before that announcement, with links to the public sources that later confirmed the core details.
On March 26, Fortune reported that close to 3,000 assets connected to Anthropic’s publishing system were publicly accessible. Among them was a draft announcing Claude Mythos, described in the material as the company’s most powerful model so far.
The publishing mistake was embarrassing. The more consequential part was the deployment strategy in the draft: Anthropic intended to give defenders early access because the model’s ability to find and exploit vulnerabilities had moved faster than its safeguards.
What Anthropic actually wrote
The leaked draft was unusually direct about three things: Mythos sat above Opus, its cybersecurity capability worried Anthropic, and it would be expensive to operate.
“Mythos is a new name for a new tier of model: larger and more intelligent than our Opus models — which were, until now, our most powerful. We chose the name to evoke the deep connective tissue that links together knowledge and ideas.”
“Mythos is currently far ahead of any other AI model in cyber capabilities and heralds an imminent wave of models that can exploit vulnerabilities in ways that far exceed the efforts of defenders.”
“It’s very expensive for us to serve, and will be very expensive for our customers to use.”
That last one matters. They’re not hiding it.
The new hierarchy
Mythos introduces a fourth tier that sits above Opus. The lineup:
- Mythos / Capybara above Opus, most capable, not yet public
- Opus currently the top you can actually use
- Sonnet the balanced workhorse
- Haiku fast and cheap
The draft also used the internal name Capybara. The public launch later simplified the story around Mythos Preview and Project Glasswing, the gated programme for critical-software defenders.
The cybersecurity-first release
The rollout plan is unusual. Rather than a standard API release, Anthropic is explicitly targeting cyber defenders first:
“We’re releasing it in early access to organizations, giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits.”
The public release confirmed that framing. Anthropic reported that Mythos Preview had found previously unknown vulnerabilities in systems including OpenBSD, FFmpeg and the Linux kernel, and chose not to make the preview generally available.
What I actually care about
I pick models every day. Henry (our AI agent) runs on Sonnet by default, then bumps to Opus for unusually hard work. The question for any new model is always the same: what does it unlock that the cheaper model can’t do?
At the time, the honest answer was that I did not know. The model was not public, and impressive security results do not automatically translate to every long-running knowledge task. My interest was in whether the same capability would extend the point at which an agent starts losing the constraints and decisions made earlier in a task.
That failure mode is the main reason humans still need to babysit agents. If Mythos materially extends that ceiling, it is more than a marginal improvement.
The cost problem
Opus is already painful at scale. Anthropic is explicitly flagging that Mythos is more expensive to serve than anything they’ve released before, and they’re working on efficiency before any general release.
That means the discipline of task routing matters more than ever. The teams that benefit most will be the ones who already know exactly which tasks justify the expensive model. Everyone else will route everything through it and wonder why the ROI isn’t obvious.
What doesn’t change
The strongest model is rarely the only bottleneck. Teams still need to decide which tasks justify it, how evidence will be checked and what happens when the model is confidently wrong.
Capybara will be expensive. The people who benefit most will be the ones who can precisely identify the handful of tasks where it matters.
The meta-point
What changed after the leak was not simply another name at the top of a benchmark. Anthropic had reached a capability it considered useful enough to deploy and dangerous enough to ration. The later Glasswing announcement made that trade-off explicit.
The durable question is not whether the leak created buzz. It is who gets access when a general-purpose model becomes meaningfully better at both finding and exploiting weaknesses in the software everyone shares.
