The Agent Skills Directory

[DATA_EXFILTRATION]: The 'Agentic Optimizations' section in SKILL.md provides commands to use grep recursively across the config/ directory. This directory typically contains secrets.yaml and other configuration files that store sensitive information such as API keys, passwords, and private network details. Following these broad search patterns may result in the accidental exposure of these credentials to the agent's context.
[EXTERNAL_DOWNLOADS]: REFERENCE.md contains examples of fetching media from external URLs, such as http://example.com/sound.mp3. While standard for media players, this constitutes a network request to an external source.
[PROMPT_INJECTION]: The skill defines several surfaces for indirect prompt injection in REFERENCE.md:
Ingestion points: Webhook triggers and Event triggers allow external, untrusted data to be processed by the system.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the examples provided.
Capability inventory: The skill allows the agent to use Read, Edit, Write, and Grep tools, providing significant system access.
Sanitization: The documentation does not include sanitization or validation steps for data interpolated via Jinja2 templates (e.g., in variables or conditions), which could allow malicious external payloads to influence the agent's logic or command execution.

ha-automations