On April 7, 2026, Anthropic officially announced the launch of Project Glasswing, a large‑scale collaborative initiative to use advanced AI capabilities to defensively protect critical global software. The project brings together more than a dozen technology and financial leaders — Amazon Web Services (AWS), Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks and others — to address new software security challenges in the AI era. The project's core driving model is Anthropic's not‑yet‑public Claude Mythos Preview, which has already discovered thousands of high‑risk zero‑day vulnerabilities in critical infrastructure, including long‑standing issues in operating systems and web browsers.
Anthropic named the initiative "Glasswing" after the glasswing butterfly, whose transparent wings symbolize how software vulnerabilities are often "invisible yet lethal." Executives including Dario Amodei emphasized that as AI coding capabilities rapidly improve, attackers might use similar models to discover and exploit vulnerabilities at scale, and Glasswing is an urgent effort to turn that capability toward defense.
Claude Mythos Preview: a cutting‑edge model so powerful it is not being publicly released
Claude Mythos Preview is Anthropic's latest frontier general model. Although it has not been formally named or publicly released, it has been made available in a limited preview to Glasswing partners. The model has achieved significant advances in autonomous vulnerability discovery, exploitation and remediation. Key highlights include:
- Vulnerability discovery: It has autonomously identified and reported thousands of high‑risk vulnerabilities across major operating systems (such as OpenBSD and the Linux kernel) and mainstream browsers. These include a 27‑year‑old remote crash vulnerability in OpenBSD, a 16‑year‑old FFmpeg bug that eluded five million automated tests, and multiple user‑to‑kernel privilege escalation chains in the Linux kernel.
- Benchmark performance: It scored 83.1% on CyberGym (vulnerability reproduction) (Claude Opus 4.6: 66.6%), reached 77.8% on SWE‑bench Pro (Opus 4.6: 53.4%), 82.0% on Terminal‑Bench 2.0, and 94.6% on GPQA Diamond. It outperforms previous generations across multimodal, agentic coding and reasoning tasks while often using significantly fewer tokens on some tasks, demonstrating efficiency.
- Autonomy: The model can complete vulnerability identification, exploit chain construction and remediation recommendations without human intervention, realizing a true "defensive hacking" capability.
For safety reasons, the model is currently limited to Glasswing partners and more than 40 critical software maintainer organizations and is not publicly available. Anthropic plans to publicly report repaired vulnerabilities, model performance and subsequent improvements within 90 days and will provide support to open‑source maintainers through a "Claude for Open Source" program.
Large‑scale collaboration and resource commitment
Anthropic is providing up to $100 million in model usage credits to Project Glasswing and has donated $4 million (including $2.5 million to Alpha‑Omega and OpenSSF projects and $1.5 million to the Apache Software Foundation). Partners will deploy the model in local environments for black‑box testing, penetration testing and endpoint hardening of first‑party software and open‑source projects.
This cross‑industry collaboration has been described as a rare "defense alliance": AI/cloud leaders such as Apple, Google and Microsoft alongside security vendors like Cisco and CrowdStrike and open‑source organizations like the Linux Foundation are jointly participating, aiming to fix vulnerabilities in critical infrastructure ahead of attackers and protect banking, healthcare, logistics, energy and government systems.
Far‑reaching impact: a turning point for software security in the AI era
The launch of Project Glasswing marks a shift from "post‑incident detection" to "proactive defense" in cybersecurity. Traditional security relies on humans or rules, while models like Mythos Preview can scan massive amounts of code at machine speed and find subtle vulnerabilities that humans may miss. This is particularly important for the open‑source ecosystem, where much critical infrastructure depends on code that has long suffered from insufficient maintenance resources.
For the industry, the initiative may accelerate an "AI security arms race": on one hand it raises overall software security, while on the other it prompts debate about the risks of spreading advanced AI capabilities. Anthropic has made clear it will not use the model for offensive purposes and intends to engage with the U.S. government and security experts to develop responsible deployment norms.
For everyday users and enterprises, Glasswing's outcomes (such as patched vulnerabilities) are expected to roll out over the coming months, indirectly improving the stability of operating systems, browsers and critical applications. The project also offers a model for AI companies on how to steer frontier capabilities toward defense rather than harm.
Challenges and outlook
Despite its promise, Glasswing faces challenges: how to balance model capability with safety boundaries, and how to prevent similar technologies from being maliciously used. Anthropic has embedded multiple safeguards in the model and pledged to release a more comprehensive "Cyber Verification Program" and invite security professionals to jointly verify the system.
Anthropic emphasized in its announcement: "The cyber threats posed by AI are evident, but if we respond correctly, we have an opportunity to build an internet and a world that are safer than before the AI era." Project Glasswing is its first large‑scale practice of that idea.
From the limited preview of Claude Mythos Preview to cross‑industry cooperation, Anthropic has again shown its emphasis on AI safety responsibility. As more organizations join, the project may become an important milestone in software security governance in the AI era and serves as a wake‑up call to the global tech ecosystem: as capabilities grow exponentially, matching defenses must be established.
▶️ Enjoy this article? Content creation is hard!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions, and shares are our greatest support! 🙏
For more AI news, visit: https://www.airosetta.com