subagents

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

• Third-party content exposure detected (low risk: 0.10). The required runtime workflow does not appear to ingest outsider-authored free text into the LLM without first selecting specific trusted sources (it primarily spawns Task members on slices using pointers/path lists, and any “web”/“repo” source-of-truth is governed by entry-skill recipes and token/structure rules rather than continuously monitored feeds).

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly references and (in one case) instructs loading/following raw GitHub documents that would be fetched at runtime to shape member prompts/output — namely https://raw.githubusercontent.com/csark0812/toolbox/main/.skeleton/references/planning/verify.md (refs/second-opinion-dispatch.md), https://raw.githubusercontent.com/csark0812/toolbox/main/.skeleton/references/verdict.md (refs/explore-escalation-dispatch.md), and https://raw.githubusercontent.com/csark0812/toolbox/main/.skeleton/references/output-schema.md (SKILL.md) which makes it high-confidence that external content could directly control prompts/output.