The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from the UniFi Protect NVR which could serve as a vector for indirect prompt injection.
Ingestion points: Tools such as protect_list_events, protect_search_detections, and protect_list_known_faces ingest user-controlled strings (e.g., face names, labels, or event metadata) into the agent's context.
Boundary markers: While the skill mandates a specific response format and separation of observations, it does not explicitly define delimiters or 'ignore' instructions for data retrieved from the NVR.
Capability inventory: The skill possesses significant control capabilities, including the ability to reboot cameras (protect_reboot_camera), disarm alarms (protect_alarm_disarm), and delete recordings or recognition data (protect_delete_recording, protect_delete_known_face).
Sanitization: There is no mention of sanitizing or escaping the data retrieved from the NVR before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill employs a dynamic invocation model using meta-tools like protect_execute and protect_tool_index. This allows the agent to discover and call any of the 63 server tools by string name at runtime. While this is a standard pattern for the underlying protocol, it facilitates dynamic behavior based on tool discovery results.

unifi-protect