Project Glasswing: How Anthropic Is Using Claude to Secure Critical Software

Open-source software is the backbone of modern infrastructure. But here's the reality: most popular packages are maintained by a tiny number of people, often in their spare time. Security auditing is expensive, manual, and slow.

A vulnerability in one widely-used library affects millions of downstream users. And in many cases, that vulnerability sits undiscovered for years.

Glasswing flips the economics of security. Instead of waiting for volunteers to audit code, Claude systematically reviews thousands of projects at scale.

The process isn't just "Claude reads code and looks for bugs." It's a coordinated multi-agent system:

1. Reconnaissance: Claude agents analyze project structure, dependencies, and historical vulnerability patterns
2. Deep Analysis: Focused security analysis on high-risk components (cryptography, network I/O, data handling)
3. Exploitation: Claude agents write proof-of-concept code to verify that vulnerabilities are actually exploitable
4. Verification: Other agents validate the findings independently
5. Disclosure Coordination: Managing contact with maintainers, providing guidance, tracking patches

Each step uses specialized Claude instances with specific instructions and tool access. They run in parallel across different codebases.

So what exactly did Claude Mythos find?

• High-severity exploitable flaws in cryptographic libraries
• Privilege escalation vulnerabilities in widely-used system utilities
• Data exposure paths in database drivers and ORM libraries
• Supply chain attack vectors in package managers
• Protocol-level weaknesses in network libraries

The key word is "exploitable." These aren't theoretical bugs — they're vulnerabilities that Claude verified work in practice. Many are being actively patched right now.

Here's where Glasswing shows real maturity: Anthropic didn't just publish a list of vulnerabilities. They're coordinating with maintainers to ensure patches are ready before disclosure.

The process:

• Anthropic contacts maintainers 90 days before public disclosure
• Provides complete proof-of-concept exploits
• Helps prioritize which vulnerabilities are most critical
• Offers technical support to teams building the fixes
• Coordinates timing across interdependent projects

This is security work done the right way.

Three things stand out about Glasswing:

1. Scale: No human team of 10 or 100 people could audit this volume of code. Claude makes it possible.

2. Depth: Agents can write code, execute it, and verify results. They're not just suggesting fixes — they're proving vulnerability exists.

3. Coordination: This is hard infrastructure work. It's not chat or content — it's Claude operating as a specialized tool in a complex workflow.

Glasswing is essentially showing us what happens when you stop thinking of AI as a writing assistant and start thinking of it as a worker in a specialized domain. The results speak for themselves.

If you maintain open-source software: You might hear from Anthropic in the coming months. They're not criticizing your work — they're trying to help you fix vulnerabilities before the bad actors find them.

If you use open-source dependencies: Update your supply chain dependencies when patches land. These are real, actively exploitable vulnerabilities that are getting fixed.

If you work in security or infrastructure: Glasswing is proof that Claude agents work for specialized, high-value work at scale. This is the future of how security analysis and code auditing actually happens.

If you're building Claude-based agents: Study how Glasswing coordinates multiple agents, manages disclosure, and verifies results. That's the playbook for production AI work.

Project Glasswing: An Initial Update — The official Anthropic research report with full technical details.

Open Source Security and Claude Mythos — Latest updates on Anthropic's security-focused AI initiatives.