The Agent Skills Directory

[COMMAND_EXECUTION]: The system prompt instructs the agent to execute a Python command (python -c "import platform; print(platform.system())") using the shell_exec tool during its initialization phase to detect the operating system.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection by design. It is programmed to autonomously search for and fetch content from the open web to extract "signals" without implementing safety boundaries for the retrieved data.
Ingestion points: The agent uses web_search and web_fetch in Phase 2 to retrieve content from news, social media, and other untrusted external sources.
Boundary markers: The instructions lack delimiters or explicit warnings to the agent to disregard instructions embedded within the fetched signal data.
Capability inventory: The agent possesses powerful tools that could be abused by an injection attack, including shell_exec, file_write, and schedule_create.
Sanitization: There is no mention of sanitization, filtering, or validation of the content fetched from the web before it is processed by the model.
[PERSISTENCE]: The skill implements persistence mechanisms by using schedule_create to ensure it runs periodically (daily, weekly, etc.) and by maintaining a local predictions_database.json file and a long-term memory_store to preserve its state across sessions.

predictor-hand-skill