huashu-nuwa

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

• Third-party content exposure detected (high risk: 0.85). 该Skill在运行时会通过“6个并行subagent”进行WebSearch并抓取公开网页/播客转录等一手或二手文本，再把这些外部来源的可读内容写入 references/research/*.md，随后在Phase 2读取这些文件进行Synthesis，从而将OUTSIDER（公共网页/转录/他人评论）自由文本喂入LLM上下文；因此存在间接提示注入风险。

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly spawns runtime subagents that fetch external research pages and media (e.g., the Tesla Master Plan URL https://www.tesla.com/secret-master-plan) and then inject those fetched sources into the Skill construction/prompt context, meaning remote content is retrieved at runtime and directly controls the agent's prompts/outputs.