Anthropic Mythos 助 AI 駭客攻擊銀行

🔑 增強重點摘要

• •Anthropic has restricted public access to the 'Claude Mythos Preview' model, citing its unprecedented ability to autonomously identify and chain together zero-day vulnerabilities in major operating systems and web browsers.
• •In response to the model's potential risks, Anthropic launched 'Project Glasswing,' a coalition of over 40 major technology and cybersecurity firms (including JPMorgan Chase, Microsoft, and Apple) granted early access to use the model for defensive patching and system hardening.
• •The model's advanced cyber capabilities emerged as a downstream consequence of general improvements in reasoning, planning, and autonomous tool use, rather than through explicit training for offensive cyber operations.

🛠️ 技術深入

• •Model Type: General-purpose frontier language model with advanced reasoning and autonomous agent capabilities.
• •Cybersecurity Capability: Capable of autonomous security research, including identifying and chaining multiple vulnerabilities into functional exploits.
• •Performance Benchmark: Demonstrated ability to discover long-standing vulnerabilities (e.g., a 27-year-old flaw in OpenBSD and a 16-year-old flaw in FFmpeg) that had previously evaded millions of automated security tests.
• •Operational Mode: Designed to function as an autonomous agent capable of working without human supervision for extended periods.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線