Project Glasswing

Anthropic’s defensive cybersecurity program through which Claude Mythos Preview is made available to a limited set of partner organizations.

Overview

• Partners maintain important software infrastructure
• Access restricted to cybersecurity use cases only
• Motivated by Claude Mythos Preview’s ability to autonomously discover and exploit zero-day vulnerabilities — capabilities valuable for defense but dangerous if broadly available
• Announced via a launch blog post accompanying the System Card

Role in Cyber Mitigations

Section 3 of the System Card (see Section 3: Cyber) confirms Project Glasswing as the primary access pathway for Claude Mythos Preview. Because of the model’s step-change in cyber capabilities — including saturating CTF benchmarks and autonomously discovering zero-days — access is restricted to vetted partners who use the model for defensive security. Probe classifiers monitor for misuse but do not block exchanges for these trusted defenders.

Context

Project Glasswing represents a novel release strategy: deploying a frontier model for a narrow, high-value use case rather than making it generally available. The same capabilities that make the model dangerous for general release (autonomous vulnerability discovery and exploitation) make it uniquely valuable for defending critical infrastructure.