白宮授權美國機構存取 Anthropic Mythos

🔑 增強重點摘要

• •The Mythos model is specifically designed for 'agentic workflows,' allowing federal agencies to automate complex multi-step tasks across secure government networks.
• •The deployment is governed by a new 'AI Security Framework' developed by CISA, which mandates human-in-the-loop oversight for all autonomous actions taken by Mythos.
• •Anthropic has implemented a specialized 'Government-Air-Gapped' instance of Mythos to ensure that federal data used for training or inference does not leak into the public model weights.

🛠️ 技術深入

• •Architecture: Utilizes a novel 'Recursive Reasoning Core' (RRC) that allows the model to self-correct during long-horizon task execution.
• •Context Window: Supports a 5-million token context window, optimized for ingesting entire federal agency document repositories.
• •Safety Mechanism: Features 'Constitutional Guardrails 2.0,' which uses a secondary, smaller model to verify the output of Mythos against federal policy constraints before execution.
• •Training Data: Fine-tuned on a curated corpus of declassified government documentation and technical manuals to improve domain-specific accuracy.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線