What is Project Glasswing?
Project Glasswing is Anthropic’s new defensive cybersecurity initiative announced on April 7–8, 2026. It uses Anthropic’s most powerful (but unreleased) AI model — Claude Mythos Preview — to proactively find and fix vulnerabilities in critical software before malicious actors can exploit them.
The name comes from the glasswing butterfly, whose transparent wings symbolize making hidden vulnerabilities visible (and thus easier to defend against) while evading harm.
Why It Was Created
Claude Mythos Preview represents a major leap in AI capabilities, especially in agentic coding and cybersecurity. It can:
- Autonomously discover zero-day vulnerabilities (including ones missed for years or decades) in major operating systems, web browsers, and foundational software.
- Chain multiple flaws to create working exploits with little to no human guidance.
- Examples include finding a 27-year-old remote crash bug in OpenBSD and a 16-year-old issue in FFmpeg that survived millions of tests.
Anthropic decided not to release Mythos publicly because its offensive cyber skills are so strong — widespread access could empower attackers and create a “cybersecurity nightmare.” Instead, they’re weaponizing it defensively through this controlled project.
Key Goals
- Secure the world’s most critical software infrastructure (OS kernels, browsers, open-source libraries, enterprise systems, etc.).
- Give defenders a head start by identifying and patching thousands of high-severity vulnerabilities.
- Share learnings across the industry and strengthen open-source security.
- Evolve cybersecurity practices for the AI era (better disclosure, patching, supply-chain security, etc.).
- Build a collaborative model so the tech industry and governments can stay ahead of AI-augmented threats.
Who’s Involved
- Core launch partners: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and Anthropic itself.
- Access extended to over 40 additional organizations responsible for critical software and infrastructure.
Anthropic is providing:
- Up to $100 million in usage credits for partners.
- $4 million in direct donations ($2.5M to Alpha-Omega/OpenSSF via Linux Foundation + $1.5M to Apache Software Foundation).
How It Works
Partners get private access to Claude Mythos Preview to scan their own systems and open-source code. The model identifies bugs, reproduces vulnerabilities, and helps develop fixes. Findings are shared responsibly (with cryptographic hashes for unpatched issues until fixes are ready). The focus is strictly defensive — terms prohibit offensive use.
Impressive Benchmark Highlights (Mythos vs. previous Claude Opus 4.6)
- CyberGym (vulnerability reproduction): 83.1% vs. 66.6%
- SWE-bench Pro: 77.8% vs. 53.4%
- Terminal-Bench 2.0: 82.0% vs. 65.4%
- SWE-bench Verified: 93.9% vs. 80.8%
It also shows big gains in general reasoning, math, and agentic tasks.
Plans & Next Steps
- Work is already underway and will continue for months.
- A public report on learnings and fixed vulnerabilities is expected in ~90 days.
- Broader collaboration on new security standards, tools, and practices.
- Anthropic invites other AI companies to join and is in discussions with the U.S. government on national security implications.
- Long-term: Develop better safeguards for future powerful models and shift cybersecurity toward AI-augmented defense at scale.
Bottom Line
Project Glasswing is a rare example of industry rivals (Apple, Google, Microsoft, etc.) teaming up with Anthropic to use a dangerously capable AI defensively — essentially racing to harden software before the same level of AI power spreads more widely and tilts the balance toward attackers.
It highlights a growing reality in AI: some capabilities are now so potent that controlled, responsible deployment is being prioritized over open release.