Project Glasswing Explained: Anthropic's Controlled AI Security Initiative

Key Points

• Project Glasswing is Anthropic's controlled-access program for Claude Mythos Preview
• 12 founding partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and Anthropic
• Over 40 additional organizations have access for critical infrastructure security
• Access is limited to defensive cybersecurity — finding and patching vulnerabilities
• No public access, API access, or commercial redistribution is permitted

What Is Project Glasswing?

Project Glasswing is the controlled-access initiative through which Anthropic makes Claude Mythos Preview available to a vetted group of organizations. Announced alongside the model itself on April 7, 2026, the program represents an unusual approach to AI deployment: rather than offering broad access and iterating on safety, Anthropic chose to limit access from the outset and expand it only as safety mechanisms are validated.

The name "Glasswing" appears to reference the glasswing butterfly (Greta oto), known for its transparent wings — perhaps a metaphor for the program's goal of making software vulnerabilities visible so they can be addressed.

The 12 Founding Partners

Each founding partner was selected because of their role in maintaining critical software infrastructure. Here is what we know about each organization's involvement:

1. Amazon Web Services (AWS) — As Anthropic's primary cloud provider and major investor, AWS is using Mythos Preview to audit the security of its cloud services and infrastructure.

2. Apple — Consumer device security is a critical concern. Apple is reportedly using the model to scan iOS, macOS, and related platforms for vulnerabilities.

3. Broadcom — The enterprise networking and semiconductor firm is focused on firmware-level security assessment for its networking products.

4. Cisco — Network infrastructure security evaluation across Cisco's routing and switching product lines.

5. CrowdStrike — The endpoint security company is using Mythos Preview for advanced threat hunting and persistent threat detection.

6. Google — Chrome browser security and Google Cloud Platform auditing are the primary use cases.

7. JPMorgan Chase — Financial infrastructure represents critical attack surfaces. JPMorgan is using the model for banking system security assessment.

8. The Linux Foundation — Open-source security is foundational. The Linux Foundation is working to identify vulnerabilities across the Linux kernel and critical open-source projects.

9. Microsoft — Windows operating system and enterprise software stack security auditing.

10. NVIDIA — GPU drivers, the CUDA framework, and AI infrastructure represent a growing attack surface that NVIDIA is working to secure.

11. Palo Alto Networks — Next-generation firewall and security platform vulnerability analysis.

12. Anthropic — Internal red-teaming, model security evaluation, and safety alignment research.

The 40+ Additional Organizations

Anthropic has stated that "over 40 additional organizations" responsible for building or maintaining critical software infrastructure have also received access. This includes organizations working on operating systems, web browsers, and key open-source libraries. The full list has not been publicly disclosed.

How Access Works

Based on available reporting, Project Glasswing operates under several constraints:

• Defensive only: Partners are authorized to use Mythos Preview exclusively for identifying and patching vulnerabilities. Offensive use is prohibited.
• Controlled environment: Access occurs through a managed infrastructure, not through standard API provisioning.
• Monitoring: Anthropic maintains oversight of how the model is used within the program.
• No redistribution: Partners cannot share access or redistribute the model's capabilities.

Comparison to Other AI Access Programs

Project Glasswing differs significantly from how other frontier AI capabilities have been deployed:

Aspect	Project Glasswing	Typical AI Model Launch
Access	12 partners + 40 orgs	Public API/web interface
Use case	Defensive security only	General purpose
Monitoring	Active oversight	Usage-based billing
Expansion	Carefully vetted	Open registration

This approach has drawn both praise and criticism. Supporters argue it demonstrates responsible AI deployment. Critics question whether a private company should decide who gets access to cybersecurity capabilities that could benefit the broader ecosystem.

For more context on the model behind the program, see our explainer page or benchmarks comparison.