The History of AI

Claude Mythos Preview and Project Glasswing

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models

Lineage
Neurons and networks
Kind
Implementation

Who

Anthropic

What happened

Anthropic announces Claude Mythos Preview, a model that according to the company finds and exploits software vulnerabilities better than almost any person, and decides not to release it. It gives access to a group of companies and organizations, Project Glasswing, so that they can fix their systems.

Why it mattered

Anthropic does not release it: it gives it to a group of organizations to fix software vulnerabilities before similar models become widespread.

The full story

On April 7, 2026, Anthropic announced a model it did not intend to release. Claude Mythos Preview is a general-purpose model, but according to the company it outperforms all but the most expert humans at finding and exploiting software vulnerabilities.

Anthropic gave examples. Mythos Preview had found thousands of serious, previously unknown vulnerabilities in the major operating systems and browsers. Among them were a 27-year-old bug in OpenBSD that allowed any machine to be crashed remotely, and a 16-year-old one in FFmpeg, in a line of code that automated tools had executed five million times without catching it. In the Linux kernel it chained several bugs together to go from an ordinary user to full control of the machine. Anthropic notified the maintainers, and those bugs were fixed.

The company feared that capabilities like these would spread beyond those who would use them carefully, and it wanted to give defenders a head start. Through Project Glasswing it gave access to the model to eleven organizations, among them Amazon Web Services, Apple, Google, Microsoft, Nvidia, and the Linux Foundation, and to more than forty that maintain critical infrastructure, with up to $100 million in usage credits and $4 million in donations to open-source security.

Anthropic announced that it would first develop safeguards capable of blocking the most dangerous uses, and that it would debut them with a later Opus model. Mozilla, for its part, fixed 271 vulnerabilities in Firefox.

Leads to

  • Claude Fable 5 and Claude Mythos 5 (June 9, 2026)

    Documented direct influence

    With Glasswing, Anthropic announced that it would build safeguards for the most dangerous uses before releasing a model of that class; Fable 5 is the first Mythos-class model with public access

  • Gemini 4 Argon (September 30, 2026)

    Inherited, with no proven direct link

    Gemini 4 Argon goes first to selected cybersecurity defenders, the access scheme Anthropic introduced with Glasswing and OpenAI followed with GPT-6 Astra

Threads that pass through here

Sources

How to cite

“Claude Mythos Preview and Project Glasswing.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/claude-mythos-preview-project-glasswing

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline