What is Project Glasswing and why it matters

In April 2026, Anthropic announced Project Glasswing — an initiative bringing together 12 key players of the tech world: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Their shared goal is to secure the planet's most critical software before attackers can exploit it.

The project's name references the butterfly Greta oto, known for its transparent wings. The metaphor is twofold: vulnerabilities hide in code in plain sight, just like transparent wings in nature, while also representing the transparency Anthropic is committed to.

At the heart of the project is Claude Mythos Preview — a not-yet-public model whose capabilities in finding and exploiting software bugs surpass all existing AI models and the vast majority of human experts. This is a model that literally changes the rules of cybersecurity.

What Mythos can do: a 27-year hidden bug and five million tests that missed it

The results of the first month, which Anthropic published on May 22, 2026 in its first update, are breathtaking. The project partners collectively found over 10,000 critical or high-severity vulnerabilities across the world's most important software.

Among the most remarkable findings are:

• A 27-year-old vulnerability in OpenBSD — an operating system considered one of the most secure in the world. The bug allowed an attacker to remotely crash any machine simply by connecting to it.
• A 16-year-old bug in FFmpeg — a library used by countless applications for video encoding. Automated testing tools hit this line of code five million times without ever detecting the flaw.
• Chaining several vulnerabilities in the Linux kernel — the model autonomously discovered and combined them to escalate from a regular user access to full control of the machine.

On the CyberGym benchmark, which simulates realistic cyberattacks, Mythos Preview achieved a score of 83.1% compared to 66.6% for the previous Claude Opus 4.6 model. On SWE-bench Pro, which tests the ability to solve real software tasks, it scored 77.8% (Opus 4.6: 53.4%).

Cloudflare, one of the partners, found 2,000 bugs in its critical systems (of which 400 were high or critical severity), with a false positive rate that was, according to their team, better than that of human testers. Mozilla discovered and fixed 271 vulnerabilities in Firefox 150 — more than ten times as many as with the previous Opus 4.6 model.

A model that must not go public

Anthropic has decided not to release Mythos Preview publicly. The reason is simple: in the wrong hands, the model's capabilities could cause catastrophic damage. Instead, the company carefully selects who gets access to the model, restricting it only to vetted partners within Project Glasswing.

This stance has led to several controversies. Anthropic denied access to the Pentagon — according to CEO Dario Amodei's statement, Defense Secretary Pete Hegseth refused to guarantee the model would not be used for mass government surveillance or fully autonomous weapons. The dispute is now being handled by an appeals court.

Similarly, Anthropic said no to China, which requested access to Mythos. Two "no's" to two superpowers — at a time when tech companies routinely navigate between commercial interests and ethics, this is an exceptional stance.

The UK's AI Security Institute confirmed that Mythos Preview is the first model to independently solve both of their cyber ranges — simulations of multi-stage cyberattacks from start to finish.

What this means for Europe and Czechia

Project Glasswing has major significance for European cybersecurity as well. The model found thousands of vulnerabilities in open-source projects on which the internet's infrastructure runs — and thus also European banks, hospitals, and energy grids. Anthropic has already reported 530 critical bugs to open-source project maintainers, but the pace of fixes is still lagging: the average time to fix a single critical bug is two weeks.

For Czech companies and institutions, the recommendation is clear: shorten security patch installation cycles to a minimum. The era when weeks or months passed between discovering a vulnerability and its exploitation is ending. With Mythos-level models, an attacker can find and exploit a hole within minutes.

The European Union is simultaneously working on implementing the AI Act, which regulates similarly high-risk models. Project Glasswing shows that even companies themselves are beginning to understand the need for self-regulation — the question remains whether it will be enough.

Against the backdrop of these events, on May 23, 2026, US President Donald Trump revoked the executive order on AI security screening under pressure from large tech companies, as reported by The Guardian. This raises further questions about who and how will oversee models with the potential to shift the global security balance.

What comes next

Anthropic plans to expand Project Glasswing to include additional partners, including US and allied governments. The company is also putting 100 million dollars in credits toward model usage into the project and 4 million dollars in direct donations to open-source organizations — of which 2.5 million dollars goes to Alpha-Omega and OpenSSF at the Linux Foundation, and 1.5 million to the Apache Software Foundation.

After the research preview ends, Mythos Preview will be available to partners at 25 dollars per million input tokens and 125 dollars per million output tokens through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

Going forward, Anthropic aims to develop sufficiently strong guardrails that will enable safe deployment of Mythos-class models even for the broader public — not just for cybersecurity, but also for other areas where similarly capable models will bring benefits.