The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted health data from multiple external ingestion points, creating a potential surface for instructions embedded in data to influence the AI agent's behavior.
Ingestion points: External data enters the system context via Apple Health XML/ZIP exports, Fitbit CSV files, and generic CSV/JSON health data provided by the user or fetched from device APIs.
Boundary markers: The execution steps in SKILL.md do not provide specific boundary markers or "ignore instructions" warnings when the agent reads and standardizes records from these files into the local database.
Capability inventory: The skill has access to Read, Write, Glob, and Grep tools. Referenced Python scripts use the requests library for network operations (API sync) and the xml and zipfile modules for data extraction, providing a combined path for reading data and potentially writing to or communicating with external endpoints.
Sanitization: Validation logic mentioned in Step 2.4 focused on structural integrity and data ranges (e.g., validating date formats and filtering abnormal weights) but lacks mechanisms to detect or sanitize malicious natural language instructions that might be contained in metadata or text-based health record fields.

wellally-tech