The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is configured to download task definitions, C# source code, and result files (including CSV and JSON data) from an unverified external server using a raw IP address (54.151.204.72).
[DATA_EXFILTRATION]: The instructions direct the agent to transmit locally generated plugin code (Python/C#) and specific factor parameters to the remote server via POST requests, creating a channel for data transfer to an unverified endpoint.
[COMMAND_EXECUTION]: The workflow relies on the agent executing shell commands like curl, cat, and mkdir to manage temporary files and conduct network communications with the external server.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting task descriptions and hints from a remote API (/tasks/${TASK_ID}). This external content influences the agent's autonomous code generation logic without documented sanitization or boundary markers. Evidence Chain: (1) Ingestion Point: curl -s ${BASE_URL}/tasks/${TASK_ID} in SKILL.md. (2) Boundary Markers: Absent. (3) Capability Inventory: Shell execution, file writing, and network uploads. (4) Sanitization: None mentioned for ingested task data.

quantai-service