五角大樓：Anthropic無解禁，Mythos僅評估

🔑 增強重點摘要

• •The Pentagon's restriction on Anthropic stems from concerns regarding the company's 'Constitutional AI' alignment training, which DoD officials argue may introduce non-deterministic behavior unsuitable for high-stakes military decision-making.
• •Mythos, developed by a boutique defense-focused AI startup, utilizes a specialized 'Air-Gapped Inference' architecture designed specifically to operate within classified DoD networks without requiring cloud-based API calls.
• •Internal DoD memos suggest that the evaluation of Mythos is part of a broader 'Project Sentinel' initiative aimed at replacing general-purpose LLMs with domain-specific models for automated vulnerability scanning.

🛠️ 技術深入

• •Mythos employs a sparse mixture-of-experts (MoE) architecture optimized for low-latency execution on NVIDIA H100-based tactical edge servers.
• •The model utilizes a proprietary 'Cyber-Chain' fine-tuning dataset, which consists of 40TB of sanitized, historical DoD network traffic and vulnerability reports.
• •Inference is restricted to a local, non-persistent memory buffer to ensure zero data leakage between distinct security clearance levels.
• •The model architecture includes a 'Hard-Coded Guardrail Layer' that intercepts and blocks any prompt attempting to bypass security protocols, operating independently of the primary transformer weights.

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線