source-driven

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

• Third-party content exposure detected (high risk: 0.85). The workflow uses WebFetch to fetch official docs pages at runtime for dependency versions (outsider-authored public web content), which can include arbitrary free text that the agent ingests into the LLM context.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly requires using WebFetch at runtime to retrieve dependency-specific documentation URLs (e.g., https://docs.example.com/v{version}/... or raw GitHub pages like https://raw.githubusercontent.com/... for each package) and then uses that fetched content to determine and drive the agent's API usage and code, so external docs fetched during runtime directly control the agent's instructions.