五角大樓與Anthropic幾近達成共識，儘管川普宣告終止

🔑 增強重點摘要

• •The 'technical misunderstandings' cited in the filing specifically involve the Pentagon's assessment of 'Model Weight Security,' where Anthropic claims the government conflated API-based access with direct access to underlying neural weights.
• •Anthropic's declarations reveal that the Department of Energy (DOE) had issued a safety clearance for the same model architecture just weeks prior, creating a direct inter-agency contradiction regarding the 'national security risk' label.
• •The filing identifies a specific March 14, 2026, meeting with the Chief Digital and Artificial Intelligence Office (CDAO) where officials reportedly confirmed that Anthropic had met 95% of the security milestones for the 'Project Aegis' deployment.

🛠️ 技術深入

• •Constitutional AI (RLAIF): Anthropic uses a 'Constitution' of principles to train models to self-critique and align with safety guidelines without human-labeled data for every edge case.
• •Model Weight Encryption: Implementation of hardware-level encryption for weights during inference in air-gapped environments to prevent physical or side-channel exfiltration.
• •Automated Red-Teaming: A proprietary framework that uses secondary AI agents to stress-test models for CBRN (Chemical, Biological, Radiological, and Nuclear) knowledge leakage.
• •Impact Level 6 (IL6) Compliance: Architecture designed for deployment within the AWS Secret Region, utilizing Nitro Enclaves for isolated compute environments.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線