Anthropic Glasswing: Securing Software for AI
Announcement of Project Glasswing — a coalition (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks) using Anthropic’s unreleased Claude Mythos Preview to find and fix software vulnerabilities.
Key Claims
- Claude Mythos Preview can surpass all but the most skilled humans at finding and exploiting software vulnerabilities
- Found thousands of zero-day vulnerabilities in every major OS and web browser, many autonomously
- Notable finds:
- 27-year-old vulnerability in OpenBSD (remotely crash any machine)
- 16-year-old vulnerability in FFmpeg (hit 5 million times by automated testing without catching)
- Multiple Linux kernel vulnerabilities chained for privilege escalation
- CyberGym benchmark: Mythos Preview 83.1% vs Opus 4.6 66.6%
- Anthropic committing 4M to open-source security organizations
- Model not generally available — access only through partners and vetted organizations
Takeaways
- This is a watershed moment: AI models are now competitive with the best human security researchers
- The offensive/defensive asymmetry is real: same capabilities that find vulns can exploit them
- The restricted access model is the source of significant controversy (see Closing of the Frontier)
- “The window between vulnerability discovery and exploitation has collapsed — what once took months now happens in minutes with AI” (CrowdStrike)
- Open-source maintainers getting access is critical — they maintain most of the world’s critical infrastructure
Connections
- The Closing of the Frontier — critique of restricted access model
- Anthropic — the company behind Glasswing
- AI Alignment — capabilities vs access tension
- AI Agent Security — broader AI security concerns