context-degradation 审计报告.

警告

• Permission surface may require sandboxing
• SKILL.md excerpt appears truncated at 'When th', making it difficult to verify the complete document is well-formed.
• The detection script explicitly relies on simulated U-shaped attention curves and heuristic token/poisoning estimates, which limits production reliability without real model internals or classifiers.
• SKILL.md does not clearly document setup, installation, or invocation details for the provided scripts and reference materials.
• No explicit limitations or safe operating boundaries section is present, though some boundaries are implied by the activation guidance.
• Quality score needs review
• Permission surface needs review: secrets or environment access, filesystem or document access
• Permission surface: secrets or environment access, filesystem or document access

方法

本报告综合公开元数据、AI 审查输出、仓库活跃度、安装就绪度、OpenAgentSkill 事件、质量评分、信任检查和 Agent 安全门槛；它不是完整的源代码安全审计。