atypica-research

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

• Third-party content exposure detected (medium risk: 0.65). The required workflow polls atypica_study_get_messages and then uses the returned message parts (including tool-requestInteraction / tool-makeStudyPlan inputs and outputs) as context for subsequent LLM/tool-calling steps, and those message parts are generated by the remote atypica service (outsider-authored free text) that the agent reads at runtime.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill makes runtime calls to the MCP endpoint https://atypica.ai/mcp/study (used by scripts/mcp-call.sh and examples) which can return persona objects including the "prompt" field that directly controls AI persona/system instructions, and the API is a required runtime dependency (API key/endpoint required).