Claude Mythos Explained: Who Has Access and Why It Matters

Claude Mythos Preview is an Anthropic model associated with advanced cybersecurity research. Project Glasswing is the initiative that gives participating organizations access for defensive work on important software.

The useful distinction is simple: Mythos is the model; Glasswing is the access and collaboration program. A news story about the model does not mean it is available in every Claude subscription or API account.

What is Claude Mythos Preview?

Anthropic introduced the preview in April 2026 and described capabilities in finding vulnerabilities and developing exploits. These capabilities can help defenders identify weaknesses, but they can also be misused.

The published research includes examples involving OpenBSD, FFmpeg, and FreeBSD. Treat those as reported findings in specific systems, not a promise that the model can successfully attack or secure any application. See Anthropic’s technical report.

A vulnerability is a weakness. An exploit uses a weakness to cause a particular effect. Finding the first does not automatically establish a reliable exploit, and generating a report does not establish that a patch is correct.

What is Project Glasswing?

Glasswing brings software vendors, infrastructure organizations, and security teams together around defensive use of the preview.

The official launch partner list includes Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. These are named on the official Glasswing page. They should not be represented as eight named partners plus four unknown slots.

The program is relevant even if your organization never uses the model: a vendor or maintainer of software you depend on may use it during security review.

Who has access?

Access has expanded since launch. On June 2, Anthropic announced plans to extend Glasswing to approximately 150 additional organizations, subject to its security requirements. That makes the initial partner count unsuitable as a permanent description of who can participate. See the expansion announcement.

Do not assume a Claude Pro, Max, or Team subscription includes Mythos Preview. Likewise, an API key by itself does not establish permission to use a restricted model. Check the official program information and the access available to your organization.

This article describes the published preview and program. It does not claim to be a live directory of all approved organizations, or to provide an access workaround.

What do the research numbers actually mean?

Anthropic reported exact agreement on severity in 89% of a sample of 198 manually reviewed vulnerability reports. That is agreement within a reviewed sample. It is not an 89% guarantee that any new model output is a real critical vulnerability.

The report also distinguishes finding a weakness from producing a working exploit. Both distinctions matter when deciding how much human review a security workflow needs. The research report is the source for these measurements; they are not independent measurements by this blog.

What should a development team do?

A practical response does not require access to Mythos. Start with the software your team already ships.

1. Identify what you run. Keep a dependency inventory, including container images and indirect dependencies.
2. Assign ownership. Each important service should have someone responsible for assessing advisories and planning updates.
3. Verify applicability. A vulnerability in a library is not automatically exploitable in every application using it. Check the affected version and deployment conditions.
4. Test the fix. Reproduce the problem in an authorized environment where possible, apply the update, and check for regressions.
5. Track deployment. A merged patch is not the same as patched production instances.

For example, if a media-processing service uses FFmpeg, record the actual version inside its deployed container. Updating a developer’s laptop does not update that container. Rebuild, deploy, and verify the running version through your normal release process.

These are engineering recommendations, not claims about a particular Glasswing partner’s internal workflow. Our deployment strategies guide explains release options.

Why human review still matters

A useful vulnerability report should identify the affected component, explain the conditions needed, provide reproducible evidence, and suggest a fix that can be tested. A confident paragraph from a model is not enough.

Anthropic’s later Glasswing update describes work with partners. Read vendor reports alongside the relevant maintainer advisories before making changes to production.

Also keep testing within systems you own or are explicitly authorized to assess. Model access does not give permission to test somebody else’s infrastructure.

Common questions

Is Mythos the same as Claude Code?

No. Claude Code is a development tool. Mythos Preview is a model discussed in Anthropic’s research and controlled-access program. A Claude Code installation does not establish Mythos access.

Does Mythos replace ordinary security tools?

No single model replaces dependency scanning, code review, testing, access controls, and incident response. Evaluate any model-assisted workflow against the task it actually performs and the evidence it produces.

Does this mean every application needs an emergency rewrite?

No. Respond to relevant, verified findings. Prioritize exposed and affected components using your normal security process rather than changing architecture solely because of a model announcement.

Where can I check availability?

Start with Anthropic’s official Glasswing announcements and your organization’s approved access channels. Avoid treating an old article, a model-name screenshot, or a third-party signup page as proof of access.

For practical controls around tools that developers can configure today, read MCP server security and protecting API keys when using coding agents.