Project Glasswing and Claude Mythos, Part 1: the initiative and what it found

An AI model has autonomously identified zero-day vulnerabilities in every major operating system and web browser, including flaws that survived decades of human review and millions of automated tests. That is not a research preview or a roadmap claim. It is a system that exists, is running now, and has restricted access while Anthropic works out what broader deployment looks like.

That system is Claude Mythos Preview, and the initiative behind it is Project Glasswing.

What is Project Glasswing?

Project Glasswing launched on April 7, 2026, as a collaborative effort between Anthropic and eleven founding partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. More than 40 additional organisations maintaining critical software infrastructure are also involved.

The stated goal is defensive: use advanced AI capabilities to find and fix vulnerabilities in critical software before attackers can exploit them. The initiative is also explicitly designed to prevent powerful AI cyber capabilities from proliferating to actors not committed to responsible deployment. The founding partners represent the organisations that run much of the world’s software infrastructure, and the Linux Foundation’s involvement signals a direct focus on the open-source layer that underpins almost all of it.

What is Claude Mythos Preview?

Mythos Preview is Anthropic’s frontier model for cybersecurity work. Its core capabilities are autonomous vulnerability discovery and exploit development: finding previously unknown flaws without human steering, and constructing the attack chains needed to demonstrate their severity.

The benchmark numbers give a sense of the capability jump. Mythos Preview scores 83.1% on cybersecurity benchmarks, compared to 66.6% for the previous-generation Claude Opus 4.6. On software engineering tasks it scores between 77.8% and 93.9% depending on the task, consistently outperforming earlier models.

What those numbers translate to in practice: Mythos Preview has identified thousands of high-severity vulnerabilities, including zero-days across every major operating system and web browser. Some of those vulnerabilities had gone undetected for decades despite continuous human security review and millions of automated test runs.

Anthropic describes this as AI reaching a level of coding capability where it can surpass all but the most skilled human security researchers at finding and exploiting software flaws. That framing matters because it sets the context for why access is restricted and why the initiative is structured as a controlled rollout rather than a standard product launch.

Who has access and how

Access to Mythos Preview is currently limited to the founding partners and selected organisations maintaining critical infrastructure. Broader availability is gated on Anthropic developing additional safeguards first, with learnings from the current rollout to be shared across the industry within 90 days of launch.

When Mythos Preview becomes more widely available, it will be priced at $25 per million input tokens and $125 per million output tokens. It will be accessible through the Claude API, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry.

Alongside the model rollout, Anthropic is committing $100 million in usage credits to support organisations doing defensive security work, and $4 million in direct donations to open-source security organisations. Open-source maintainers are specifically named as a target beneficiary, which reflects the reality that the open-source layer is both the most widely used and historically among the least resourced when it comes to systematic security review.

What this means for the code you ship

Three things worth taking from this announcement as a software engineer or tech lead.

First, the capability threshold has shifted. Code you write and ship will increasingly be reviewed by systems that can find vulnerabilities that eluded both expert human reviewers and extensive automated testing. That raises the floor on what “thorough security review” means, and it raises it in a direction that is automated and scalable.

Second, open-source security is getting direct investment. The $4 million in donations and explicit focus on open-source maintainers addresses a gap that has been a persistent source of risk across the industry. The dependencies your applications rely on are part of what this initiative is aimed at.

Third, the 90-day publication commitment means this is not a closed experiment. What Mythos finds, what the effective safeguard patterns look like, and what the false positive and false negative rates look like in practice will enter the public conversation within the year. That is worth tracking if your team is making decisions about AI-assisted security tooling.