dialogue-gym

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

• Third-party content exposure detected (medium risk: 0.30). SKILL.md’s runtime workflow ingests outsider-provided free text because the user pastes their own dialogue “attempt” (and intake framing) for the coach to read and score, without requiring the agent to first select a specific trusted item.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly states it connects "over the MCP connector" to read a user's Map and fetch scenes at runtime—behavior that would pull external content (https://writecalliope.ink/docs/mcp) into the agent's context and thus directly influence its prompts/instructions.