HA Integration Dev
Pass
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The documentation for the conversation agent template provides instructions for users to download and install Ollama, an established service for running local Large Language Models.
- [PROMPT_INJECTION]: The conversation agent template creates a functional surface for indirect prompt injection by ingesting untrusted device states into an LLM prompt to drive service execution.
- Ingestion points: The LLM agent in
templates/conversation-agent/conversation_agent.pyretrieves the names and states of controllable devices (lights, switches, climate, etc.) and includes them in its system context. - Boundary markers: The system prompt uses structural headers to delimit device data, but it lacks specific instructions for the LLM to ignore potential commands hidden within device names or state attributes.
- Capability inventory: The agent has the capability to execute Home Assistant service calls (e.g.,
turn_on,set_temperature) via thehass.services.async_callmethod. - Sanitization: The agent validates the LLM's response using regular expressions and JSON parsing to ensure it adheres to a strict action schema before calling any services.
Audit Metadata