Anthropic, a leading AI developer, has unveiled its latest language model, dubbed ‘Claude Mythos Preview’. This advanced model displays a remarkable ability to autonomously identify and exploit vulnerabilities in software. Unlike traditional systems, Claude Mythos is said to surpass the capabilities of all but the top human security experts.
Decision Against Public Release
Due to its potential risks, Anthropic has decided against a public release of Claude Mythos. The model’s extraordinary capabilities raise concerns about its misuse by irresponsible actors. The company fears that should these abilities fall into the wrong hands, the consequences could threaten economic and public safety, as well as national security.
Internal Tests Reveal Critical Vulnerabilities
During internal evaluations, Claude Mythos Preview identified numerous zero-day vulnerabilities, which are security flaws unknown to developers. The findings include:
- A 27-year-old vulnerability in OpenBSD, allowing a network connection to crash affected machines.
- A 16-year-old flaw in FFmpeg that previously went unnoticed despite automated testing.
- Multiple interrelated vulnerabilities in the Linux kernel that could grant complete control to an attacker.
These issues have been reported to the respective software maintainers and have since been addressed. For additional vulnerabilities, Anthropic has only released cryptographic hashes and plans to disclose more details post-resolution.
Project Glasswing: A Defensive Initiative
To leverage the model’s defensive capabilities, Anthropic launched ‘Project Glasswing’. This initiative aims to utilize Claude Mythos Preview for enhancing cybersecurity measures while sharing insights across the industry.
Key Partners Involved
Prominent organizations involved in Project Glasswing include:
- Amazon Web Services
- Apple
- Broadcom
- Cisco
- CrowdStrike
- Microsoft
- NVIDIA
- Palo Alto Networks
In total, over 40 organizations, dedicated to developing or maintaining critical software infrastructure, will access the model for vulnerability assessments.
Financial Commitments and Outlook
Anthropic has allocated up to $100 million for usage credits of Claude Mythos Preview under Project Glasswing. An additional $4 million will be distributed as donations to open-source security organizations.
The company emphasizes that the battle for global cybersecurity could take years. With rapid advancements in frontier AI, Anthropic calls for immediate action from cyber defenders. The message is clear: collaboration between AI developers, software firms, security researchers, and governments is essential to tackle these complex challenges effectively.
