Anthropic Unveils Security Model Mythos

Anthropic has made Claude Mythos available in a limited release through Project Glasswing, aimed at security research partners rather than the general public. The model's cybersecurity research capabilities are extremely powerful, at a level capable of discovering high-risk vulnerabilities in operating systems and web browsers.

Key Performance and Features:

• Vulnerability Chaining: Rather than a single vulnerability, it can sequentially combine 3-5 vulnerabilities to create sophisticated exploits.
• Autonomous Exploit Development: While the existing Claude 4.6 Opus had a near-0% success rate in autonomous exploit development, Mythos Preview recorded a very high success rate. (For example, in a Firefox JS engine vulnerability experiment, it succeeded 181 times out of hundreds of attempts.)
• Real-World Cases: It autonomously performs tasks such as browser sandbox escape using JIT heap spraying, Linux local privilege escalation, and FreeBSD RCE exploits.

Security experts are warning that AI-generated security reports have moved beyond the previous low-quality 'AI slop' stage and are now entering a stage where they represent extremely sophisticated, real threats.