Anthropic announced Project Glasswing on April 7, 2026, a cybersecurity initiative centered around Claude Mythos Preview, a frontier AI model with exceptional autonomous hacking capabilities that the company will deliberately not make publicly available. The model achieved state-of-the-art performance on coding benchmarks while autonomously discovering thousands of zero-day vulnerabilities across major operating systems and web browsers.
Claude Mythos Preview Achieves State-of-the-Art Coding Performance
Claude Mythos Preview scored 93.9% on SWE-Bench Verified and 87.3% on SWE-Bench Multilingual, representing the highest performance on these industry-standard coding benchmarks. The model demonstrated capabilities beyond traditional code generation, autonomously identifying and exploiting a 17-year-old remote code execution vulnerability in FreeBSD that allows root access on machines running NFS. Mythos Preview can chain multiple exploits and perform multi-step attack sequences without human guidance.
Anthropic Restricts Model Access Due to Dual-Use Risks
Unlike typical AI releases, Anthropic explicitly stated they "do not plan to make Claude Mythos Preview generally available" due to significant dual-use risks. Instead, the company granted early access to 12 partner organizations including Amazon/AWS, Apple, Broadcom, Cisco, CrowdStrike, Linux Foundation, Microsoft, NVIDIA, Google, Palo Alto Networks, and JPMorgan Chase, with plans to expand to 40 organizations. Anthropic is in discussions with the US government about both offensive and defensive cyber capabilities.
$104 Million Committed to Software Security Initiatives
Project Glasswing includes substantial financial commitments to strengthen critical software infrastructure:
- $100 million in model usage credits for partner organizations
- $2.5 million donated to Alpha-Omega and OpenSSF through the Linux Foundation
- $1.5 million to the Apache Software Foundation
These investments focus on identifying and patching vulnerabilities in widely-used open-source software that underpins global digital infrastructure.
Industry and Community Response
The announcement reached #1 on Hacker News with 860 points and 384 comments, with three related posts (the main announcement, system card PDF, and red team assessment) simultaneously hitting the front page. The release represents a significant departure from the AI industry's typical approach of broad model availability, signaling recognition that certain AI capabilities require new governance frameworks.
Key Takeaways
- Claude Mythos Preview achieved 93.9% on SWE-Bench Verified and autonomously discovered thousands of zero-day vulnerabilities across major operating systems and browsers
- Anthropic will not make the model publicly available, instead restricting access to 40 partner organizations due to dual-use cybersecurity risks
- The company committed $104 million total, including $100 million in model credits and $4 million to open-source security initiatives
- 12 major tech companies and financial institutions received early access, including Apple, Microsoft, Google, NVIDIA, and JPMorgan Chase
- This marks the first major frontier AI model deliberately restricted from public release due to security concerns