# What is Claude Mythos Preview, and why is Anthropic keeping it for cyber defenders?

Date: 2026-04-08 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-claude-mythos-preview-and-project-glasswing/

**Answer:** Claude Mythos Preview is a new, unreleased Anthropic model that is so good at finding and exploiting software flaws that Anthropic is giving it only to a group of security partners, through an effort called Project Glasswing announced on April 7, 2026. The goal is to let defenders find and patch weaknesses in widely used software, from operating systems to browsers, before attackers gain similar tools.

Yesterday, Tuesday, April 7, 2026, Anthropic announced a model it does not plan to release to the public. **Claude Mythos Preview** is a general-purpose model, like other Claude models, but Anthropic says its coding skill has reached the point where it can find and exploit security holes in software better than all but the most expert people.

Instead of a normal launch, Anthropic created **Project Glasswing**, a partnership that gives the model to companies and open-source groups that maintain critical software, so they can use it for defense. Newton Cheng, who leads cyber work on Anthropic's Frontier Red Team, told VentureBeat the company does not plan general availability because of these cyber capabilities.

## What Anthropic announced

Twelve organizations are launch partners. Besides Anthropic, they are big technology companies (Microsoft, Google, Apple, Amazon Web Services, NVIDIA and Broadcom), security specialists (Cisco, CrowdStrike and Palo Alto Networks), the bank JPMorganChase, and the Linux Foundation, which supports open-source software. More than 40 further organizations that build or maintain important software also get access.

| Commitment | Detail |
|---|---|
| Usage credits | Up to $100 million of Mythos Preview use during the research preview |
| Donations | $4 million in all: $2.5 million for the Alpha-Omega and OpenSSF security projects (given via the Linux Foundation), plus $1.5 million for the Apache Software Foundation |
| Price after credits | $25 per million input tokens, $125 per million output tokens, for participants only |
| Public report | Lessons and fixed flaws to be shared within 90 days |

Several partners say they have been testing the model for weeks. The name comes from the glasswing butterfly, whose transparent wings make it hard to spot, a fitting image for old bugs that sit unnoticed in familiar code.

## What the model found

A **vulnerability** is a flaw in software that an attacker could use, for example to crash a system or take control of it. An **exploit** is the method that actually uses the flaw. A **zero-day** is a vulnerability the software's makers did not know about, so no fix exists yet.

Anthropic says Mythos Preview has found thousands of serious zero-days, including flaws in each of the main operating systems and web browsers, and that it found nearly all of them without human guidance. Three examples have already been fixed:

1. **OpenBSD**, an operating system known for its security: a 27-year-old flaw that let anyone crash a machine remotely just by connecting to it.
2. **FFmpeg**, video software used inside countless apps: a 16-year-old flaw in one line that automated tests had already exercised five million times without catching it.
3. **Linux**, which runs most of the world's servers: several flaws chained together to turn an ordinary user account into full control of the machine.

For the many flaws not yet patched, Anthropic is publishing only a **cryptographic hash** of each report: a short fingerprint that proves later that it knew the details on this date, without revealing them now.

Anthropic's own benchmarks compare the model with Claude Opus 4.6, its best released model:

| Test | Mythos Preview | Opus 4.6 |
|---|---|---|
| CyberGym (reproducing known security bugs) | 83.1% | 66.6% |
| SWE-bench Verified (fixing real code issues) | 93.9% | 80.8% |
| SWE-bench Pro (harder coding tasks) | 77.8% | 53.4% |

## How the defense is supposed to work

The plan is a head start. Software has always had hidden bugs, and finding the dangerous ones used to need rare human experts. If AI makes that search cheap, attackers will eventually benefit too. Glasswing tries to let the people who maintain the software go first.

Responsible disclosure is part of the design. VentureBeat reports that Anthropic sends its findings through paid human reviewers before passing them to maintainers, agrees a pace with each project so volunteers are not flooded, and attaches a suggested patch labeled as model-written when it can. After a fix is available, it generally waits 45 days before publishing full technical details, so users have time to update.

Here is the kind of instruction a partner's security team might give a model like this, inside their own systems:

```prompt
Review the attached network-parsing module from our router firmware.
List any memory-safety or input-validation weaknesses, rank them by
severity, explain how each could be triggered, and propose a minimal
patch with a test. Do not run anything against live devices.
```

> **Try it:** Pick an app you use every day and check whether automatic updates are on, on your phone and computer. Patches for flaws like these reach you only if you install them.

## Why it matters

Anthropic is choosing to keep its most capable model away from the public because of what it can do. The company argues the risk is close: similar abilities are likely to spread before long, possibly to people who will not use them safely. It plans to test new safeguards on an upcoming Claude Opus model before considering wider use of models of this class.

The announcement also invites scrutiny. VentureBeat points out that the model's existence first came out through a leaked draft blog post in late March, followed days later by a separate Claude Code packaging error, awkward events for a company asking to be trusted with such a tool. Anthropic calls both human publishing mistakes, not breaches of its core systems.

## What does Project Glasswing mean for ordinary computer users?

Mostly, it means more security updates. If partners find and fix flaws in operating systems, browsers and common software, the fixes will arrive through the normal update buttons on your devices. Installing them promptly is the single most useful response.

You cannot use Mythos Preview yourself, and you do not need to. The released Claude models remain available for everyday tasks, including ordinary help with code.

It is also a sign of a broader change. Several partners warn that AI will speed up attacks as well as defenses, so habits like strong unique passwords, two-step sign-in and regular updates matter more, not less. Anthropic has promised a public report within 90 days, which should show how much the project has actually fixed.

## Sources

- Project Glasswing: Securing critical software for the AI era: https://www.anthropic.com/glasswing
- Anthropic says its most powerful AI cyber model is too dangerous to release publicly — so it built Project Glasswing: https://venturebeat.com/data/anthropic-says-its-most-powerful-ai-cyber-model-is-too-dangerous-to-release