Claude Mythos wasn't announced. It leaked.
A data leak revealed the model's existence before Anthropic said anything publicly. Then the company confirmed it, described it as "the most capable we've built to date," and made clear they have no plans for a general release — at least not yet.
What followed was Project Glasswing: a controlled testing initiative with some of the biggest names in tech, using Mythos Preview to find and fix critical security vulnerabilities before the model reaches the public.
This is not a typical AI launch. It's worth understanding what's actually happening here.
What Mythos Preview can do
The cybersecurity numbers are striking. In the weeks since Project Glasswing began, Mythos Preview has:
- Identified thousands of zero-day vulnerabilities across major operating systems and web browsers
- Written functional exploits in hours that expert penetration testers estimated would have taken them weeks
- Autonomously found and exploited a 17-year-old remote code execution vulnerability in FreeBSD — one that allows root access on any machine running NFS
These aren't benchmark results. These are outcomes from actual red-team work against production software.
Anthropic's framing is that Mythos represents "a step change" — not an incremental improvement on the Claude 4 line, but a qualitatively different level of agentic coding and reasoning capability. The cybersecurity results are the clearest public evidence of what that means in practice.
Project Glasswing
Anthropic didn't open Mythos to the public. Instead they formed a consortium of companies to use it for defensive security work under controlled conditions.
The current Project Glasswing partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and Nvidia.
The logic is straightforward: a model capable of finding critical zero-days and writing working exploits at this speed is dual-use by definition. Anthropic wants to build defences ahead of broader deployment — getting defenders equipped before the same capabilities become accessible to anyone else.
Why it isn't shipping yet
Anthropic has been explicit: Mythos Preview will not be made generally available in its current form.
The company says it wants "new safeguards in place" before deploying Mythos-class capabilities at scale. The cybersecurity preview is part of how those safeguards get built — using the model to harden systems, then assessing whether the defensive gains outpace the offensive risks.
This is a more cautious deployment approach than anything Anthropic has done with the Claude 4 series. It suggests Mythos is genuinely different in kind from what they've released before, not just a stronger version of the same thing.
What we don't know yet
Anthropic hasn't published a technical report. There's no confirmed parameter count, no public benchmark suite, and no announced API pricing. What we have is the leak, Anthropic's confirmation, and the Project Glasswing results.
The capability claims around cybersecurity are specific and verifiable — the companies involved aren't going to stake their security programs on marketing. But how Mythos performs outside the security domain, what the context window looks like, and what the eventual deployment model will be are all still unknown.
What to expect
Anthropic has said they "eventually want to safely deploy Mythos-class models at scale." That implies a general release is planned, but with conditions attached that don't exist yet.
Given the Project Glasswing timeline started in early April, and assuming the consortium's defensive work needs months to produce meaningful results, a public release in 2026 is plausible but not certain.
When it does ship, it will almost certainly be behind usage policies tighter than anything currently on the Claude platform. The cybersecurity capabilities alone make that inevitable.
For now, Mythos remains one of the most capable AI models ever built — and one that isn't available to use. Project Glasswing is the closest anyone outside Anthropic is getting to it, and the results coming out of that program are making clear why the company is being careful.
Sources:
- Exclusive: Anthropic is testing 'Mythos,' its 'most powerful AI model ever developed' — Fortune
- Anthropic debuts preview of powerful new AI model Mythos in new cybersecurity initiative — TechCrunch
- Project Glasswing: Anthropic announces big tech consortium to test Claude Mythos — IT Pro
- Anthropic's Unreleased Claude Mythos Might Be The Most Advanced AI Model Yet — PYMNTS