arkweb-architect

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). 该技能在 Phase 2 Step 2.0/2.2/2.4 将用户需求（{user_requirement}、{tech_keywords_from_requirement}）以及由此产生的“知识库证据包”（包含原始检索内容）直接注入到多个 subagent 的 task 文本中（如专家团、code-analysis、brainstorm），从而让外部用户可通过其提交的自由文本影响 LLM 实际读取的运行时输入。

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly instructs runtime retrieval of external DeepWiki pages (e.g., https://deepwiki.com/openharmony/arkui_ace_engine) which are persisted into a "知识库证据包" and then injected into subagent task prompts, so fetched content would directly control model prompts at runtime.