Anthropic Shelves Mythos Over Hacking Risks

🔑 Enhanced Key Takeaways

• •Anthropic has restricted access to the Mythos model to a select group of approximately 40 cybersecurity and technology partners under an initiative called 'Project Glasswing' to focus on defensive patching rather than public deployment.
• •Technical testing revealed that Mythos achieved a 72% success rate in identifying and creating working exploits for software vulnerabilities, a massive leap from the near-0% success rate of previous models like Opus 4.6.
• •The model has demonstrated the ability to autonomously discover 'zero-day' vulnerabilities in legacy and heavily audited codebases, including a 27-year-old bug in OpenBSD and a 16-year-old flaw in FFmpeg, which had previously evaded automated detection tools.

🛠️ Technical Deep Dive

• •Model Architecture: Part of the Claude family, specifically optimized for autonomous vulnerability research and exploit chain development.
• •Performance Metrics: Demonstrated 83.1% success rate on 'CyberGym' benchmarks (testing against real open-source codebases) compared to 66.6% for Opus 4.6.
• •Exploit Generation: Capable of autonomous chaining of Linux kernel issues to achieve full machine control and splitting complex ROP (Return-Oriented Programming) chains over multiple packets.
• •Testing Methodology: Utilizes a scaffold that isolates the project-under-testing and its source code, allowing the model to focus on specific files to identify remote code execution (RCE) vulnerabilities.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline