evaluate

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

• Third-party content exposure detected (medium risk: 0.30). Runtime path: the workflow first resolves a provided URL and then clones the resulting git repository (WebFetch/mcp__fetch__fetch for docs, then survey.sh does git clone --depth 1 into /tmp/evaluate/<slug>/ and reads files to extract license/manifests/tree/LOC and passes subsets to subagents for read/grep), which is outsider-authored free text from the specified repository.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill runs a shallow git clone of an arbitrary repository URL ("https://github.com/owner/repo(.git)" in survey.sh) at runtime and then has subagents read that fetched code to drive prompts/findings, so remote repo content can directly control agent behavior.