HA Integration Dev
Pass
Audited by Gen Agent Trust Hub on May 15, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The LLM Conversation Agent template in
templates/conversation-agent/conversation_agent.pyfacilitates an indirect prompt injection surface.\n - Ingestion points: The agent ingests external user input (
user_input.text) and Home Assistant entity states (viaself.hass.states.async_all()) into a dynamically constructed system prompt.\n - Boundary markers: The prompt template utilizes section headers such as
AVAILABLE DEVICESandRULES, but lacks randomized or unique delimiters to isolate untrusted data from instructions.\n - Capability inventory: The integration is capable of executing arbitrary Home Assistant service calls via
self.hass.services.async_callbased on the LLM output.\n - Sanitization: The agent utilizes regex patterns and standard JSON deserialization to extract intended actions, providing a layer of validation between LLM responses and system execution.\n- [EXTERNAL_DOWNLOADS]: Documentation in
templates/conversation-agent/README.mdprovides an installation command for the Ollama service using a shell script fetched fromhttps://ollama.com/install.sh. This reference points to a well-known service for executing large language models locally.
Audit Metadata